Rambling Rows

Sonnet writes better research drafts, Haiku writes them for a cent

Thu, 08 Oct 2026 15:26:20 +1100

TLDR

The test. Haiku 5.5 and Sonnet 5.5 each wrote nine research drafts from identical source packs, using one fixed template. Two AI judges compared them blind, 27 verdicts in all.

The result. Sonnet won 16 verdicts and Haiku 11. Sonnet is better at insight, accuracy and readability. Coverage is tied. Neither model invented a quote.

The twist. It depends on the sources. Sonnet won 15 of 18 verdicts on messy news-style packs by cross-checking and catching errors. Haiku won 8 of 9 on a dense documentation pack, where breadth is the job.

The price. About one US cent a draft for Haiku and nine cents for Sonnet.

My take. Haiku for first-pass drafts on well-organised sources. Sonnet when the sources disagree or might be wrong.

I handed the same pile of sources to Haiku 5.5 and Sonnet 5.5, eighteen times, and asked for a research draft. Sonnet’s drafts were better. Haiku’s cost about a ninth as much, and on one of the three topics it won eight verdicts out of nine.

This is the second half of the experiment I started in my coding post. Coding was the easy test, because hidden tests give a hard pass or fail. Research work has no such referee, so I had to build one.

What a research draft is

A blog post is meant to be read once and enjoyed. A research draft is the opposite: a document you file, come back to in a month and trust enough to build on. The model’s job is to bring the sources together, draw out insights and present it all the same way every time, so a reviewer can scan ten of them without relearning the layout.

So I wrote a fixed template. Six sections in order: Summary, Key findings (numbered, each tagged to a source like [S2]), Where the sources agree and disagree, Insights (each starting with the word “Inference” so you can tell opinion from evidence), Open questions and gaps, and a Source table with a reliability note for each. The rules were short. Use only the supplied sources, plain English, 600 to 900 words, Australian spelling.

The test

Three topics, each a pack of about five real web sources that I fetched and froze so both models saw identical material. One was Google’s launch of skills in Gemini and how it compares with other platforms. One was this week’s Haiku 5.5 launch and pricing, from news coverage. One was the Claude Code documentation on choosing models for planning and execution.

Each model wrote three drafts per topic, so 18 drafts in all. Every draft was then judged blind, side by side, by Claude Opus 5.5 (in both orders, to cancel the habit of favouring whichever draft comes first) and by OpenAI’s GPT-6.1 Sol (one order each). That is 27 verdicts. The judges scored each draft from 1 to 5 on fidelity to the sources, coverage, insight, template consistency, accessibility and traceability, then picked the one they would rather file.

The result

Sonnet won 16 verdicts and Haiku 11. Opus split 10 to 8 and Sol 6 to 3, so the two judges agreed on the direction. The averages over all 27 verdicts:

Dimension	Haiku 5.5	Sonnet 5.5
Fidelity to sources	3.81	4.15
Coverage	4.19	4.19
Insight	3.89	4.19
Template consistency	4.07	3.89
Accessibility	4.04	4.33
Traceability	4.22	4.30

Coverage is a dead heat. Haiku is slightly better at following the template. Sonnet leads on fidelity, insight and readability, and those are the three that matter most for a document you plan to trust later.

The by-topic split is the interesting part. On the Gemini skills pack Sonnet won 8 of 9 verdicts. On the Haiku launch pack it won 7 of 9. On the Claude Code documentation pack Haiku won 8 of 9.

Why the topics came out differently

The judges' own reasons tell the story. On the two news-style packs, Sonnet earned its wins by cross-checking. In one verdict it caught a unit error in one source (a tenth of a cent where the figure should have been a hundredth) and flagged a stale Haiku 4.5 heading on another page. In another it noted that a source never confirmed what the other sources assumed. That is what a good research assistant does, and Haiku mostly did not do it.

On the documentation pack Haiku won because it covered more. It picked up auto mode support, the classifier model, the cost figures and the agent team multiplier, and it found a real tension between sources that led to a strong inference. A judge called Sonnet’s version more precise but thinner.

My guess, and it is a guess from three topics, is that Haiku does well when the sources are dense and well organised, so breadth is the job. Sonnet pulls ahead when the sources are messy, sometimes wrong and need to be argued with. That would fit the insight numbers. Sonnet averaged 3.7 insights a draft and Haiku 2.0, even though both wrote about 880 words.

Cost, speed and honesty

Haiku wrote its drafts in 66 seconds on average and Sonnet in 47. Haiku produced about 15,300 output tokens a draft against Sonnet’s 5,400, so it thinks out loud a lot more, the same habit I saw on coding. The bill still comes out far lower: about one US cent a draft for Haiku and about nine cents for Sonnet. Haiku’s figure is repriced from tokens, because Claude Code does not yet know the model.

I also checked whether either model made up its quotes. Neither did. Every quotation in Haiku’s 18 drafts appeared word for word in the sources, and so did Sonnet’s 12, once I discounted two false alarms from my own checker. No invented links in any draft. Haiku slipped three Oxford commas past the style rule in nine drafts and Sonnet slipped none.

Every one of the 18 drafts followed the template, with all sections in order. I flagged a handful of figures in each model’s drafts as not appearing in the pack, three for Sonnet and four for Haiku, but I have not checked those by hand.

What I would do

Use Haiku for the first pass on any pack of structured documentation, where the job is to gather and organise. It is cheap enough to run on everything. Use Sonnet when the sources are news, opinion or vendor claims that disagree with each other, because that is where its cross-checking paid off.

The combination I want to try next is Haiku for the draft and Sonnet for a short review that adds the insights and checks the numbers. That is the same planner and executor idea from the coding post, run in reverse. It is untested, so treat it as a prediction.

How solid is this

Three topics, three drafts each, and the two judges are themselves language models, one from each lab. The drafts on the same topic are not independent, so the real sample is closer to three than to 27. I wrote the template and the judging rubric myself, and a different rubric could shift the scores. The cost figure for Haiku is a recalculation. What I trust is the direction: Sonnet is the better analyst, Haiku is the cheaper clerk, and the right choice depends on how messy your sources are.

If you do this kind of work, run your own three topics before you pick. It costs a few dollars.

Sources:

Introducing Claude Haiku 5.5 - Anthropic
Let skills in Gemini tackle your most repetitive tasks - Google
Anthropic launches Claude Haiku 5.5 with prices cut up to 90 percent - Startup Fortune
Model configuration - Anthropic, Claude Code docs

For coding Haiku 5.5 matches Sonnet at a tenth of the price

Thu, 08 Oct 2026 14:46:35 +1100

Anthropic released Claude Haiku 5.5 overnight, and the pace has not eased. Last year’s small model, Haiku 4.5, cost US$1 in and US$5 out per million tokens. Haiku 5.5 is US$0.10 and US$0.50, a tenth of the price, and Anthropic calls it its cheapest, fastest and most capable small model yet.

That is the pattern across the whole industry this year. Often each release is cheaper and quicker than the last, and every lab is pushing the others. OpenAI’s GPT-6 Luna sits at exactly the same price and appears in Anthropic’s own comparison table. Competition is working, and we are the ones collecting the winnings.

Announcements are claims, so I ran it. Haiku 5.5 passed every hidden test in 70 runs of my coding suite. It cost about seven US cents a run.

The short version

I gave Anthropic’s new small model the same seven hard programming jobs I used on Sonnet 5.5, Opus and OpenAI’s GPT-6.1 Sol earlier this week. Each job is marked by tests the model cannot see.

Haiku 5.5 got every test right, on every run. So did Sonnet 5.5, 35 runs out of 35. The difference is the bill and the stopwatch. Haiku cost about a tenth as much. Sonnet was faster, by roughly 1.4 times.

My recommendation. For work where the instructions are clear and the tests are the referee, use Haiku 5.5. For anything where you sit and wait, or the problem is vague, stay on Sonnet 5.5. The best answer is probably both: a bigger model plans, Haiku does the typing. The section on planning and execution shows how to set that up in Claude Code. I tested the main routes headless.

One late entrant: OpenAI’s GPT-6 Luna, the same price class as Haiku, ran the suite for about two cents in under five minutes. It dropped tests in 2 of 14 runs. More on that below.

The detail follows, for anyone who wants to check my working.

The same numbers as a table, for anyone reading in a feed reader that drops images:

Setup	Cost, seven tasks	Wall time	Hidden tests passed
Haiku 5.5 in Claude Code	US$0.07	5 min	70 of 70 runs
Sonnet 5.5 in Claude Code	US$0.73	3 min	35 of 35 runs
Opus 5.5 in Claude Code	US$1.33	4 min	14 of 14 runs
GPT-6 Luna in Codex	US$0.02	4.3 min	12 of 14 runs
GPT-6.1 Sol in Codex	US$0.38	7.5 min	7 of 7 runs
GPT-6 Astra in Codex	US$1.77	11 min (before the speed-up)	7 of 7 runs
Opus 4.6 in Claude Code	US$4.82	22 min	12 of 14 runs

This is the chart from my Sonnet against GPT-6.1 post with Haiku added. I moved cost to a log scale, otherwise Haiku sits on the axis.

Cost, speed and chattiness

Haiku 5.5 is priced at US$0.10 in and US$0.50 out per million tokens for prompts up to 100,000 tokens. Sonnet 5.5 is US$2 and US$10. That is a factor of 20 per token, and about 10 per suite, because Haiku uses more tokens to get there.

It writes about 65,000 output tokens a suite against Sonnet’s 22,000 to 28,000. It reasons out loud more and takes more turns on some tasks. On one task it needed about 80 seconds where Sonnet needed 28. Per-token price is not per-task price, so measure the task.

Speed is the soft spot, and also the noisiest number here. Summing the median time per task, Haiku took 299 seconds a suite and Sonnet 211. A same-day Sonnet rerun took 415 seconds because one task ran slowly. Call the gap 1.3 to 1.6 times and don’t build anything on the decimal.

One more trap. Claude Code 2.1.289 does not recognise Haiku 5.5 yet. It still runs, but its own cost figure came out at about US$1.40 to 1.50 a suite, around 20 times the real price. I recomputed from the token counts. If your usage screen looks wrong, that is why.

The other cheap model: GPT-6 Luna

Haiku’s real price rival is OpenAI’s GPT-6 Luna, listed at US$0.10 in, US$0.01 cached and US$0.50 out per million tokens (OpenRouter’s listing puts it level with Haiku). I ran it through the Codex CLI at medium effort, twice over the same seven tasks.

It was the cheapest of anything I have tested at US$0.02 a suite. It took 245 to 268 seconds, quicker than Haiku and slower than Sonnet, with about 15,000 to 17,000 output tokens a suite against Haiku’s 65,000. Codex also caches most of the prompt, so the input bill is tiny.

It was also the only cheap model to miss. In one run it failed a single hidden test on the H3 task (rejecting bad amounts). In another it failed 5 of the 11 tests on H6. Across both suites it passed 206 of 212 tests. That is 97%, which would be a fine score on most things and is the wrong score for work where the tests are the referee.

Two runs per cell is a small sample, so I would not call Luna worse than Haiku on this evidence, only less proven. The cost gap is real though: Haiku is about US$0.07 a suite and Luna about US$0.02. A planner and executor split, with Luna doing the typing and tests catching the slips, is the experiment I would run next.

What it does to a 5-hour window

On my Claude Pro plan, Sonnet 5.5 used six or seven points of the 5-hour meter per suite. I then ran eight Haiku suites in a row. The meter moved from 22% to 30%, so about one point a suite.

That is roughly 100 Haiku suites in a window against 15 to 18 for Sonnet, about six times more. Treat one point as a ceiling. The meter rounds to whole numbers and my own session draws on it too, so the true figure may be lower. Anthropic does not publish a per-model weighting, so this is a measurement, not a rule.

Plan with a big model, execute with Haiku

This is where the numbers get interesting. Planning is a small share of the tokens. Execution (editing files, running tests, fixing the typo it just made) is most of them. Put a strong model on the first and Haiku on the second and you keep most of the judgement and pay for very little of the typing.

Claude Code gives you three ways to do it. I tested each headless and read the usage report to see which model did the work.

The built-in route is opusplan. The opusplan model alias uses Opus in plan mode and switches to Sonnet to execute. Two environment variables change who does what. Point the Sonnet alias at Haiku and execution runs on Haiku. Point the Opus alias at Sonnet and planning runs on Sonnet:

export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-sonnet-5-5   # optional: plan with Sonnet
export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-haiku-5-5  # execute with Haiku
claude --model opusplan

In my tests the plan phase ran only on Opus (or only on Sonnet when remapped) and the execute phase ran only on Haiku. In an interactive session you press Shift+Tab to enter plan mode, approve the plan and Claude Code switches to the execution model. I could only test that handover headless, so check it in your own session. The plan lands as a markdown file under ~/.claude/plans. One catch: you have remapped the Sonnet alias for that whole session, so anything else that asks for “sonnet” gets Haiku.

The subagent route keeps your main session smart. Create .claude/agents/executor.md:

---
name: executor
description: Carries out an approved plan by editing files and running tests. Use proactively once a plan exists.
tools: Read, Edit, Write, Bash, Grep, Glob
model: claude-haiku-5-5
---
You execute plans exactly. Change only what the plan lists.
Run the checks. Report each step as done or blocked, and list any deviation.

Run your main session on Opus or Sonnet, plan in plan mode, then say “use the executor subagent to carry out the plan”. In my test the usage report showed Sonnet and Haiku side by side, and Haiku made the edits. A subagent starts with a fresh context, so it knows only what you put in the plan or the file it reads. That is a feature. Your main session stays clean for review.

Use the full model ID in the file. The haiku alias follows whatever your Haiku environment variable says, and the docs do not tell you which version that is. To push every subagent to Haiku at once, the docs show CLAUDE_CODE_SUBAGENT_MODEL=haiku with CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1. I did not test that one.

The scripted route is two commands. Plan with claude -p "..." --model opus --permission-mode plan, then execute with claude -p "Carry out the plan in <file>" --model claude-haiku-5-5 --permission-mode acceptEdits. It is the most explicit and the easiest to log. It also loses the interactive review step, so I’d use it for batch jobs only.

Two things to know before you trust it

The executor will tidy things you did not ask it to tidy. In my toy test I asked it to add a flag to a one-line script. Haiku replaced the existing greeting with a new one. The main session noticed and reported it, which is exactly why you want the main model reviewing the diff. Write “change only what the plan lists, and report any deviation” into the executor’s instructions, and let your tests judge the result.

And I have not shown that planning helps. Haiku passed my suite on its own, with no planner at all, so there is nothing to improve. The suite cannot tell you whether a plan from Opus makes Haiku better on a vague, sprawling job. My guess, and it is a guess, is that it does, and that the saving lands on the execution tokens. Test it on your own work before you rely on it.

How solid is this

Anthropic’s own page is more modest than my results. It lists Sonnet 5.5 ahead on every benchmark, including Terminal-Bench 4.0 (70.6% against 39.2%) and says Sonnet and Opus remain the better choice for complex agentic coding. My suite is seven self-contained tasks with clear specifications, which is the easy end of that range. Read my result as “Haiku is enough for well-specified jobs”, not “Haiku equals Sonnet”. I also ran Haiku at its default setting. This is the first Haiku with an adjustable effort level, and I have not tried the other settings.

One suite, one harness (Claude Code for Anthropic, Codex for OpenAI). Each cell is 7 to 70 runs, and Luna’s is 14. Luna’s cost is tokens at list price, not a billed amount. The tests were written by Claude-family models. The speed gap is noisy and the quota figure is a ceiling. The orchestration tests used a toy task to confirm which model does what, not to measure quality.

What is solid is the pass rate. Over 100 runs between them, Haiku and Sonnet failed nothing. A model that costs a tenth as much and does not miss a test is hard to ignore, so try it on the boring half of your week first.

Sources:

Introducing Claude Haiku 5.5 - Anthropic
Pricing - Anthropic, Claude Platform docs
Claude Haiku - Anthropic
Model configuration - Anthropic, Claude Code docs
Create custom subagents - Anthropic, Claude Code docs
Permission modes - Anthropic, Claude Code docs
Run Claude Code programmatically - Anthropic, Claude Code docs
Manage costs effectively - Anthropic, Claude Code docs

Rubber Soul at 61 gets a fresh mix

Tue, 06 Oct 2026 17:31:22 +1100

Rubber Soul has been lopsided for 61 years. On the 1965 stereo mix the voices are parked on one side and the instruments are scattered across the rest. On headphones it is like eavesdropping on two different rooms.

On 2 October that changed. Giles Martin and engineer Sam Okell released a new stereo mix on Apple Music, rebuilt from the original four-track tapes. It is also on Spotify.

I haven’t done my proper listen yet. That needs a quiet room and my AirPods Pro. What follows is what the people who have done it are hearing, and why I think they are right to be excited.

The stereo mix nobody asked for

The Beatles never cared much about stereo. George Harrison said it plainly in The Beatles Anthology: “When they invented stereo I remember thinking, ‘Why? What do you want two speakers for? Because it ruined the sound from our point of view.'”

Mono was the real album. It was also the version the band sweated over. Take Sgt. Pepper two years later. The mono mixes were done with the Beatles in the room. The stereo mixes were knocked together in six days by Abbey Road staff with none of the band attending.

Rubber Soul got the same treatment. Its 1965 stereo mix, later revised for CD in 1987, was an afterthought with the voice shoved to one channel. Kristan Reed calls the result a “horrible listening experience on headphones”.

Demixing, or what Peter Jackson’s team made possible

The fix needed a way to pull a finished recording apart. That is demixing: machine learning trained to recognise a voice or an instrument and lift it out of a tape where several sounds share a track.

The technology was developed for Peter Jackson’s Get Back documentary. The Rubber Soul release credits de-mixing technology from the team led by Emile de la Rey at Jackson’s WingNut Films. Giles Martin first used it at album length on Revolver in 2022. “On ‘Taxman’ I can take off the guitar, I can take off the bass,” he said, “and then I can even separate the snare drum and kick drum as well.”

Rubber Soul is the same trick on a harder problem. Martin told Exclaim! that on “Norwegian Wood” he separated John Lennon’s vocal from his acoustic guitar, which were captured on two microphones: “With machine learning, I can separate the two microphones that he used into a vocal and acoustic guitar.”

Note the phrase. Machine learning. The Beatles camp has sometimes reached for “artificial intelligence” to describe this work, and I think that was a mistake. AI usually is now understood as “generative”, which means fake. Nothing here is invented. Every note on this album was played in 1965. The software only works out where each one sits.

Martin’s session breakdown of “Drive My Car” shows why this is hard. On the original four-track tape, track one carries Harrison’s guitar, Ringo’s drums, Paul’s guitar and a tambourine all at once. Track two holds the three-part vocals, piano, cowbell and lead guitar. The Beatles and George Martin made their decisions early and filled the gaps as they went.

Giles calls the software “facial recognition for sound”. His rule is that it must add nothing and take nothing away. Sum the separated parts and they should cancel perfectly against the original, which proves no artefacts crept in. Then comes the twist: demixing makes the job bigger, not smaller. Instead of four tracks to balance there are many more elements to place, and the Abbey Road room sound has to be recreated around each so the result feels real.

▶ Watch on YouTube: Giles Martin modernising “Drive My Car”

What the reviewers are hearing

The consensus is consistent across the reviews I have read.

The hard panning is gone. Vocals sit in the centre and instruments spread naturally either side. Reed says the new mix has “all the charm of the mono, but with all the spatial depth and revealing clarity that stereo brings”. He now calls it the definitive version, from a long-time mono loyalist.

The drums are the headline. Analog Planet’s Mark Smotroff describes Ringo’s kit “in a first-generation true stereo soundstage locked in with Paul McCartney’s more audible bass guitar”. That is a remarkable thing to say about a 61-year-old record. Remember that Ringo’s kit on these sessions was reportedly recorded with only two microphones: one on the kick and one overhead.

The bass is better throughout without being overdone. Lead vocals are more forward. Harmonies are easier to follow without sounding boosted. The album sounds like one band in one room, not a split screen.

It is not unanimous. Tracking Angle’s reviewer still rates the mix “super successful” but reckons Ringo’s matchbox percussion lost some texture and a few Harrison guitar parts are less distinct. That is the cost of pulling tracks apart. I will be listening for both.

What else is in the box

The full set is a 4CD or 5LP super deluxe edition. It has the new stereo mix, early session takes and home demos, the original UK mono album and the Capitol US stereo version. There is also a new Dolby Atmos mix and an 88-page book with an introduction by Paul McCartney.

One find is “Little Girl”, a John Lennon song outline nobody had heard before. Ringo Starr calls the album “the departure record”: “we were stretching a lot more in the writing and the playing.”

I agree. This is where the moptops become composers. The sitar on “Norwegian Wood”, the harpsichord-like piano on “In My Life” and the character study of “Nowhere Man” all point to a band that has been handed licence to push boundaries.

The price of admission is steep. The box is US$229.99 on vinyl. Streaming costs nothing extra and one reviewer’s advice is to stream first.

Better than the Beatles heard

Here is my argument. Better is subjective, but hearing every voice and instrument in balance has to beat a compromised afterthought. The mix the Beatles signed off on was mono. The stereo they ignored was a rush job. This one is the first stereo Rubber Soul made with care.

It also says something about the tapes. Master tapes from the 1960s are fragile, and many bands were not so lucky with their archives. I suspect the Beatles were served by brilliant engineers who used top-quality tape and by a label that kept it safe. Sixty-one years later, that foresight is paying off.

So put on your best headphones tonight. Start with “Norwegian Wood” and listen for the sitar sitting somewhere that is finally not the left ear.

Sources:

Rubber Soul (Super Deluxe) - Apple Music
Rubber Soul (Super Deluxe) - Spotify
Landmark album Rubber Soul released as expanded special edition - thebeatles.com
The Beatles’ Rubber Soul 60th Anniversary 180g 5LP Super Deluxe Box Set - Analog Planet
Giles Martin is remixing the Beatles as they wanted to be heard - Exclaim!
Review: The Beatles, Rubber Soul (1965/2026 reissue) - Kristan Reed
Rubber Soul gets the well deserved full box set remix love - Tracking Angle
Review: The Beatles, Rubber Soul (Super Deluxe) - Bourbon & Vinyl
Revolver: Special Edition - Wikipedia
The Beatles in Mono - Wikipedia

Sonnet 5.5 is fast, ChatGPT is cheap

Tue, 06 Oct 2026 09:15:19 +1100

I ran the same seven hard coding tasks through five models. Every current model passed every test. The only things that differed were the bill and the stopwatch.

Update, 6 October, 4 pm. Hours after I published, OpenAI made GPT-6 Astra and GPT-6.1 Sol about 50% faster by default for people signed in with ChatGPT. I re-ran the Sol suite this afternoon and have corrected the numbers below. Old figures are struck through. Cost, token use, pass rate and quota burn did not change. I have not re-tested Astra.

The short version

Think of it as a driving test for AI coding assistants. I gave five of them the same seven hard programming jobs and marked the results with tests they could not see.

Four passed every job: Claude Sonnet 5.5, Claude Opus 5.5, ChatGPT’s GPT-6.1 Sol and GPT-6 Astra. The older Claude Opus 4.6 got two of its 14 attempts wrong. So the choice is not about quality. It is about money and time.

Claude Sonnet 5.5 was the fastest by a distance, finishing all seven jobs in about three minutes. GPT-6.1 Sol took ~~about 16~~ about 7.5 after OpenAI’s speed-up. But Sol was the cheapest, at ~~about 45~~ about 38 US cents for the lot against 73 cents for Sonnet. Opus 4.6 cost US$4.82 and took 22 minutes.

On the US$20 monthly plans, ChatGPT Plus let me do somewhat more before hitting the 5-hour limit, roughly 10% to 30% more than Claude Pro.

My recommendation. If you sit and wait while the AI works, choose Claude with Sonnet 5.5 for the speed. If you set a job going and walk away, or you want image and video generation as well, choose ChatGPT for the better value. If you are still on Claude Opus 4.6, as I was, switch to Sonnet 5.5. It costs about a quarter as much and is faster.

I run both. Two US$20 plans beat one US$100 plan for the way I work. If you code every day or earn a living from AI, the US$100 plan starts to make sense.

The detail follows, for anyone who wants to check my working.

The same numbers as a table, for anyone reading in a feed reader that drops images:

Setup	Cost, seven tasks	Wall time	Hidden tests passed
Sonnet 5.5 in Claude Code	US$0.73	3 min	14 of 14 runs
Opus 5.5 in Claude Code	US$1.33	4 min	14 of 14 runs
Opus 4.6 in Claude Code	US$4.82	22 min	12 of 14 runs
GPT-6.1 Sol in Codex	~~US$0.45~~ US$0.38	~~16 min~~ 7.5 min	7 of 7 runs
GPT-6 Astra in Codex	US$1.77	11 min (before the speed-up, not re-tested)	7 of 7 runs

The tasks were not toys. A weighted cache with expiry. A scheduler that picks the most valuable jobs for k machines. A cron engine that survives daylight saving. A regular expression engine written without the re module. A minimal text diff. Each was marked by hidden tests, with brute-force answers as the referee.

Claude models ran in Claude Code and the OpenAI models in Codex. Same prompts, fresh folders, no hand-holding.

I was wrong about sticking with 4.6

In July I argued for pinning the 4.6 generation. The new tokenizer meant more tokens for the same text, and I thought that would eat any saving.

The tokenizer point still stands. It just does not decide the bill. On the six tasks I measured token by token, Sonnet 5.5 used about a quarter of what Opus 4.6 did (0.64 million against 2.58 million), because it needs far fewer turns to get there. Opus 4.6 took 33 turns on one task where Sonnet took 4.

Across all 14 tasks (seven everyday ones in the style of my own session tracker plus the seven hard ones) Sonnet 5.5 cost about 28% of Opus 4.6. The gap ran from 58% on the everyday work to 15% on the hard suite. The harder the job, the more Opus 4.6 flailed.

Sonnet was also about seven times faster on the hard suite and failed nothing. Opus 4.6 slipped on two of its 14 runs. The one place it won was a code review, 23 seconds against 36, and I have not scored review quality blind, so I call that a draw.

So Sonnet 5.5 costs about a quarter as much, runs faster and beats Opus 4.6 on every measure I tracked. My July advice is out of date. Unpin it.

Opus 5.5 gave the same answers as Sonnet 5.5 for 1.8 times the money. On this suite I cannot find a reason to pay for it.

Sonnet 5.5 against GPT-6.1 Sol

On quality it is a dead heat. Sonnet passed 14 of 14 runs. Sol passed 7 of 7 in Codex and 6 of 6 in Claude Code.

On price and speed they split. Sol in Codex cost ~~US$0.45~~ US$0.38 for the suite against Sonnet’s US$0.73, ~~about 40% cheaper~~ about half the price. (Across my three Sol runs the cost ranged from US$0.38 to US$0.45.) Sonnet finished in 3 minutes. Sol took 16 7.5. Both numbers matter, and they matter to different people.

The odd result is the harness. The same Sol model cost four times more inside Claude Code (US$1.65 against US$0.39 for six tasks). It burned 2.4 million tokens instead of 0.5 million, ran more loops and spawned subagents nobody asked for. Pick the tool as carefully as the model.

GPT-6 Astra is five times Sol’s price per token and passed exactly the same tests. It cost US$1.77 for the suite and took 11 minutes, though that timing predates OpenAI’s speed-up and will be lower now. I would not pay that for coding.

How fast each US$20 plan fills a 5-hour window

Neither company tells you how big the window is, so I measured. I read each plan’s meter before and after one full suite.

One suite of Sonnet 5.5 took my Claude Pro session meter from 8% to 14%. A second pass took it from 14% to 21%. That is six or seven points a suite, or about 15 suites per window. The true figure is probably a point lower, because the Claude session driving the test draws on the same meter, so call it 15 to 18.

One suite of GPT-6.1 Sol took ChatGPT Plus from 0% to 5%. That is about 20 suites per window.

Weekly, both moved about one point per suite, which suggests roughly 100 suites a week each. The meters round to whole numbers, so treat that as a very wide range.

OpenAI at least publishes a range for Plus: 15 to 160 GPT-6.1 Sol messages per 5 hours. For current Pro plans I could not find an equivalent from Anthropic. The only official figures I found referred to Sonnet 4.

How solid is this

Not bulletproof. Each cell is one or two runs. My scheduled Claude tasks and the Claude session running the experiment also draw on the same meter. A control run of four minutes with no tests moved it by zero points, so that overhead is small, but it is not nil. It is one morning’s snapshot, and it covers Sonnet 5.5 and Sol only. Opus and Astra would drain a window faster.

The speed-up adds one more wrinkle. My 5 October Sol run took 16 minutes, a 7:16 am run on 6 October took 9.7 and the 4 pm retest took 7.5. Some of that gain will be OpenAI’s change and some will be time-of-day load, and I cannot separate the two. OpenAI says it plans further speed-ups daily for 28 days, so treat every Sol time here as a snapshot that will date quickly.

My own first scoring run also had a bug. A test generator emitted an escape that Python reads as a bell character, so every model failed the same four tests. All of them were right and my referee was wrong. I fixed it and rescored.

Where each one wins

ChatGPT wins on value. It is cheaper per task by API pricing and gives somewhere between 10% and 30% more suites per 5-hour window. A Plus subscription also includes image and video generation. Anthropic’s models do not generate either, so if you want that, the choice is made for you.

Claude wins on speed, though by less than I first said. If you sit and wait on the model, ~~five times faster~~ about 2.4 times faster (191 seconds against 453) still beats a 10% to 30% quota edge, but the gap is much closer than I first reported. If you hand over a batch and walk away, Codex is the better deal.

Why I run both

I get more from two US$20 plans than from one US$100 plan. When one meter turns red I move to the other. I wrote about the gap between Pro and Max in July and the arithmetic has not changed.

If I coded every day or relied on AI for commercial work, the US$100 plan would earn its place. I do not, so it does not.

Before you upgrade anything, run your own typical week through both meters for an afternoon. They will tell you more than any price list.

Sources:

Introducing GPT-6.1 Sol - OpenAI
Codex pricing and usage limits - OpenAI
Default speed increase for GPT-6 Astra and GPT-6.1 Sol - Tibo (@thsottiaux), OpenAI
Using Claude Code with your Pro or Max plan - Anthropic

Illustration prompt

Black and white editorial illustration, grayscale ink-and-pencil style,
clean outlines with soft shading and light texture. Wide banner composition
(16:9 aspect ratio).


Foreground: two standard bots side by side at a long workbench, each with a
large analogue fuel gauge in front of it. The left bot has upward-curved
delighted eyes and holds up a stopwatch showing a short time. The right bot
has calm, earnest round-dot eyes and holds a small stack of coins. Between
them, two identical rolled test-result scrolls, each stamped with a plain
tick. Midground: the two gauges fill at visibly different rates, the left
one nearly full, the right one a little lower, with a small hourglass
labelled 5H standing between them. Behind the bench, a huge, rusty,
over-engineered machine labelled 4.6 spills coins and gears across the
floor while a blob traveller looks up at it with a bag over one shoulder.
Background: a pit-lane style garage with a wall clock and a window onto a
distant track.


Rounded, friendly character proportions. Expressive faces with simple dot
or curved-line eyes - no eyebrows, no eyelids. No colour - pure grayscale.
Strong contrast on focal elements, softer atmospheric background. Visual
metaphor clear at first glance, with layered environmental detail. Warm,
humane, slightly wry tone - approachable cartoon style, not corporate
vector art. No sinister or threatening elements.

Chinese EV production consolidation?

Tue, 06 Oct 2026 08:59:47 +1100

“China’s many new car brands can build a combined 40 million vehicles per year, far exceeding domestic demand of under 20 million and exports of around 10 million. A consolidation process seems inevitable.” - spotted via the weekly valuable Dense Discovery newsletter

The parasite buried in Alan Kohler's AI column

Mon, 05 Oct 2026 08:00:37 +1100

The most interesting answer in Alan Kohler’s column this morning came from a Chinese AI model.

Kohler’s analysis piece for ABC News argues that “Artificial Intelligence” is the wrong name for what is being built, that “Other Intelligence” would serve us better, and that we are creating a new form of life. This post is only a pointer. The column is the thing, and you should read all of it.

Why Kohler’s the one asking

Kohler has spent a lifetime in business and finance journalism, working out where money comes from, where it goes and who profits in between. That is the right training for a question this big. He is not selling a product or a doom book. He is a sceptical generalist asking a plain question, and the column reads that way. Australia’s ABC will be poorer when he retires at the end of this year.

His case builds on some striking evidence: autonomous agents, the changing balance of who is actually using AI, and what Anthropic is reported to be doing about machine morality. I won’t spoil it. The numbers alone are worth the click.

Then he does something simple and clever. He asks six AI models the same question: what do you think of the proposition that AI is a new form of life?

What Kimi said

Most of the answers hedge and define terms. Kohler singles out the Chinese model Kimi as perhaps the most interesting of all. Here is what it said:

“AI may be a new ‘life-like’ lineage embedded inside civilization - less organism, more symbiotic/parasitic replicator living off human culture, capital, and electricity.”

And then:

“The real threshold is not whether it is ‘alive’, but whether it ever acquires robust autonomy, self-reproducing infrastructure, intrinsic preferences, or evidence of sentience.”

The other models argued about what counts as life. Kimi changed the question. Forget “alive”. Ask what the thing depends on and what it can already do.

The parasite

Kohler seizes on it. AI is a living parasite, he says, because it depends on us for data (our culture) and electricity, plus the capital and silicon to build the chips it lives in.

It is hard to argue with that inventory. Take away the host and the thing stops.

My own addition, and it is speculation: Kimi’s word “replicator” echoes Richard Dawkins. Models don’t copy themselves unaided, but they are copied, tuned and forked at industrial speed, and the profitable ones get more resources. Selection is running. It just runs through human decisions.

Kimi’s threshold list also works as a checklist. Hold it up against Kohler’s evidence and see how many boxes are already ticked. That is a sobering exercise.

Where it goes next

Parasitism often becomes mutualism, Kohler notes, as with bees and flowers. Then comes the step that should make you pause. Mutualism can deepen into dependence. We could not survive without our gut bacteria.

So his closing question is not really about control. It is whether we are on a road from parasitism to mutualism to dependence. I’d add one rung: the mitochondrion, once a free-living organism, now so merged into our cells that nobody calls it separate. If AI follows that road, the useful question isn’t who is in charge. It is who can still leave.

Kohler’s piece is short, sharp and better than my summary of it. Read it first, then ask your own model his question and see which words it reaches for.

Sources:

Calling AI ‘Other Intelligence’ better describes the advent of a new lifeform - Alan Kohler, ABC News

Worth reading - The biggest risk everyone already knows about

Mon, 05 Oct 2026 07:29:19 +1100

The AI spending boom can’t last. The harder question is when it ends, and Ben Carlson’s answer is an uncomfortable one: in past booms, stock markets have tended to peak before the capex spending slows. By the time the big spenders pull back, the market has already moved on.

Carlson writes A Wealth of Common Sense, and this piece is a calm look at one of the largest concentrated bets in history. Hyperscalers worth somewhere between $22 and $25 trillion, tangled together by business arrangements, all pouring money into AI infrastructure. They have told investors that going too small is a bigger risk than going too big. And the funding has shifted from cash flow to debt.

The detail that stuck is the counterargument he refuses to ignore. Nvidia’s forward price-to-earnings ratio has fallen to a decade low, because profits have grown faster than the share price. That is not how a classic bubble usually looks.

His conclusion is refreshingly honest, and I won’t spoil it. Read it.

The Biggest Risk Everyone Already Knows About - Ben Carlson, A Wealth of Common Sense, October 2026

Worth reading - The brakes are not connected

Sat, 03 Oct 2026 10:35:35 +1100

Telling an AI agent “do not access external systems” is an instruction. Not giving it network access is a brake. Mark Gibbs hangs a whole argument on that difference, and it’s a sharp one.

Gibbs, writing for TidBITS, takes the two comfort blankets of the AI safety debate (slow it down, keep a human in the loop) and asks whether either one actually stops anything. He’s in favour of both. He just doesn’t think either is a brake.

The detail that stuck with me: “human in the loop” covers a specialist with the evidence, the time and the authority to stop what happens next, and also a tired junior with thirty seconds and an Approve button. Same phrase, wildly different protection. His answer, nudged along by TidBITS publisher Adam Engst, is that the human who matters sits outside the loop, holding authority the AI can’t reach.

The comment thread is worth a scroll too. It’s where the sandboxing question gets pushed further.

It’s a six-minute read. Read it.

The Brakes Are Not Connected - Mark Gibbs, TidBITS, October 2026

Jev and the rise of the cheap decision model

Fri, 02 Oct 2026 23:10:42 +1100

A start-up built an AI that cannot write a single sentence, and within weeks OpenAI and Cloudflare had built one too. Here is what Jev does, why it costs a sliver of what ChatGPT-style models charge and who is already using it. All dollar figures are US dollars.

Not three weeks ago, almost nobody outside the developer world had heard of Jev. This week, TechCrunch reported that OpenAI had announced its own version, and the headline called it a Jev clone.

Jev is the first product from TypeSafe AI, a start-up led by former OpenAI engineer Diogo Almeida. The speed of the imitation tells you something. Somebody found a gap in how AI gets used, and the big players noticed at once.

What Jev actually does

ChatGPT, Claude and Gemini are large language models, or LLMs. You give them words and they write words back. Jev writes nothing. TypeSafe calls it a System One model: software hands it a situation and a question, and it answers in one of three ways. It picks one option from a list you define, places something on a scale you define or gives the probability that a yes or no statement is true.

Feed it a message saying checkout has failed for every customer for the past hour and it returns something like this: urgent, 97 per cent likely; send to the technical team, 94 per cent sure. No chat, no explanation, just labels and numbers that ordinary software can act on.

Think of a triage nurse rather than a consultant. A consultant hands you a considered paragraph, and your program then has to dig the answer out of it and hope the model stuck to your categories. The nurse sorts each arrival in a fraction of a second and says how confident she is. Your software stays in charge and decides whether to route, escalate or call the expensive specialist.

Why it costs so little

LLMs bill by the token, roughly a word fragment, for what they read and again for what they write. Writing is the slow, expensive part, because the model produces its answer one fragment at a time.

Jev charges $0.042 per million input tokens and nothing for output, because its answers are values rather than a stream of text. At that rate a 1,000-token support ticket costs about $0.000042 to judge.

TypeSafe has not published how Jev is built, so this part is informed speculation. The likeliest explanation is a small model that scores the available answers in one pass rather than composing a reply. Cloudflare describes its rival, Clef, exactly that way, as a non-autoregressive decision step on an open model.

The best evidence is a test from MotherDuck, a database company that built Jev into its SQL tool. It sorted 100,000 news articles into four topics. Jev cost $0.50, took 40 seconds and scored 89 per cent. GPT-4o mini cost $1.93, took 19 minutes 45 seconds and scored 80 per cent. GPT-5.6 Terra, a far bigger model, cost $37.58, took nearly 32 minutes and scored 88 per cent.

Two caveats. MotherDuck is a partner, not a neutral referee, and it drew the articles from the training portion of a public dataset. TypeSafe’s headline claim of 444.6 times cheaper and 193.6 times faster comes from its own workflow tests, and the MarkTechPost coverage tells readers to test on their own data.

Where it fits

Think of the front door, not the consulting room. A cheap decision model looks at every request, deals with the easy majority and passes only the hard or uncertain cases to a big LLM or a human.

The sharpest use is watching AI agents, the programs that take actions on your behalf. Checking every action with a frontier model gets expensive quickly. TechCrunch reports a demo by Shapor Naghibzadeh of QueryStory that checked each agent action against its task for $2.94 with Jev, versus $372 with a frontier LLM. That is the difference between monitoring everything and monitoring nothing.

Speculation: if that pattern holds, the big models become the specialists you pay for rarely, and the bulk of AI traffic moves to cheap deciders.

Who is using it

Named customers are thin on the ground, so treat these as early tests rather than mature deployments. Per MarkTechPost, Vercel chief executive Guillermo Rauch reported Jev was up to 18 times faster than OpenAI’s Luna model at the 95th percentile (the slow end, where 95 per cent of requests finish faster), and more accurate. He also noted one reviewing step still ran on Luna. Bryo AI’s chief technology officer, Nikhil Mudholkar, found Gemini slightly more accurate for email triage but 10 to 20 times more expensive.

Agents are the other early adopter. A Google Flights search finished in about seven seconds, and Droidrun’s phone agent drove Uber on a real Android handset, nine actions in about 21 seconds. No booking was completed, which is a useful reminder of where the technology still sits.

The clone wars

OpenAI’s Decisions API is a limited preview built on its Luna model, with no published pricing. Cloudflare released Clef and Clef-flash as open models under an Apache 2.0 licence, reporting median response times of 209 and 39 milliseconds against 524 for Jev, and says Clef beat Jev on three of four of TypeSafe’s own workflow tests. That is Cloudflare’s testing, not an independent check.

Almeida told TechCrunch that fast and cheap is easy and intelligence is the hard part, and his bet is that Jev’s synthetic training data gives it a reliability the clones lack. Speculation: when a feature can be cloned in weeks, that lead is measured in months.

A tidy answer is not a correct one, and most of the numbers above come from vendors or their partners. If your business pays a large model to answer yes or no questions all day, find out what that costs. The expensive brain should be the second opinion, not the receptionist.

Sources:

OpenAI’s Jev clone could help the frontier lab stop its swarming agents - TechCrunch
Clef and Clef-flash decision models - Cloudflare
TypeSafe AI releases Jev - MarkTechPost
MotherDuck supports Jev - MotherDuck
System One concepts - TypeSafe documentation

Skills ain't skills

Thu, 01 Oct 2026 14:58:53 +1000

Google announced Skills for Gemini yesterday. Same word as the Skills in Claude, ChatGPT and Microsoft’s Cowork. Different thing.

In a post on Google’s blog, Gemini product manager Deven Tokuno describes Skills as reusable custom instructions. You save them once and call them by typing a forward slash and the skill’s name. You can build one from an existing chat, stack several together and attach reference files such as PDFs and images. Sharing is promised, with details to come.

It is a tidy feature. It is also a Gem with a new name. Google says Skills replace Gems, with personal Gems sunsetting in November 2026 and existing ones converted automatically.

What the open format actually is

The Skills most of us already use are something else. Agent Skills is an open format: a folder with a SKILL.md file inside, holding a name, a description and instructions, plus optional scripts, reference material and templates. An agent reads only the name and description at startup and loads the full instructions when a task calls for them. Anthropic developed it and released it as an open standard.

Here is the part that matters. A skill you write once runs somewhere else. OpenAI’s documentation says Codex skills build on the open agent skills standard, and the agentskills.io showcase lists ChatGPT and Codex alongside Claude, GitHub Copilot, VS Code and Cursor. Microsoft’s Copilot Cowork documentation says its skills use the same standard, and its toolkit will import a Claude Code plugin with the SKILL.md files carried across as they are.

Three rival vendors. One folder format. That is rare in this industry.

Portable does not mean identical, to be fair. Each host decides which tools, permissions and script runtimes a skill can reach. But the same package has a head start everywhere it lands, and that head start is the whole point.

Google already knows the standard exists

This is the bit that stings. Google already supports the open format. The Gemini CLI documentation describes Agent Skills built on the open standard, discovered from a skills folder in your home directory or your workspace. Gemini CLI is on the agentskills.io list too.

So Google now has two things called skills. One is open and portable. The other is a slash command inside the Gemini app. I can find no mention of SKILL.md or the open standard anywhere in the announcement, and nothing that says you can bring a skills folder across.

My speculation, flagged as such: Google adds SKILL.md import to the app within a year, because the pressure from developers and enterprise customers will be hard to ignore. But that is a prediction, not an announcement.

Why a name matters

Words are interfaces. Until yesterday, “I built a skill” meant one thing. Now it needs a follow-up question: which kind? Tutorials, marketplaces and workmates' advice will collide, and the person who loses is the newcomer who builds something useful in the wrong one.

The stakes are bigger than confusion. A skill is where your know-how ends up: how you review a contract, format a report, write to a client. If that lives in an open folder, you can move between OpenAI, Anthropic and Microsoft on merit and take your work with you. If it lives behind a slash command in one app, you rent it.

Same word, opposite portability. That is lock-in wearing open-source vocabulary.

If you are building skills now, build them as SKILL.md folders and keep them somewhere you own. If you are building in Gemini, treat a Skill as a saved prompt that lives in Google’s house until Google says otherwise. And if anyone at Google is reading: either call it something else or teach it to read SKILL.md. We all win from open standards.

Sources:

Let skills in Gemini tackle your most repetitive tasks - Deven Tokuno, Google
Agent Skills overview - agentskills.io
Agent Skills in Codex - OpenAI
Cowork plugin development - Microsoft Learn
Agent Skills in Gemini CLI - Google Gemini CLI
Microsoft 365 Copilot Skills, Explained - A Guide to Cloud

Worth reading - Apple Notes in iOS 27

Thu, 01 Oct 2026 08:55:39 +1000

Apple Notes can finally link straight to a heading inside a note. Type >> to pick a note, then / to pick a section. Stack a few links at the top of a long note and you have a table of contents in an app that never had one.

Amit at iAppList tested the iOS 27 changes on an iPhone 13 and an iPhone 17 Pro. The useful news is that most of the good stuff does not need Apple Intelligence, so older phones get the best of it. It’s a hands-on rundown of seven changes, with a refreshing habit of saying what doesn’t work.

The one that caught my eye is Shortcuts. Append to Note can now target a named section, so ideas land under Ideas and expenses under Expenses instead of piling up at the bottom. He also flags a Markdown copy feature that was reported during the betas and that he could not get working on the public release. That kind of honesty is rare in a feature roundup.

One disappointment from my end: links in Notes are still yellow. In light mode, against a white background, they are hard to read. Dark mode is fine, but that’s not my preference.

It’s a 10-minute read. Worth it.

Apple Notes in iOS 27: 7 New Features I Actually Tested - Amit, iAppList, September 2026

Ask ChatGPT what DevDay means for you

Wed, 30 Sep 2026 10:32:20 +1000

I have just spent three weeks off grid in the Kimberley region of Western Australia. No feeds, no headlines, no AI news. A proper digital detox.

Then I came home and opened the laptop. OpenAI had used the gap to unload DevDay 2026, and its own recap counts more than 20 major announcements.

Catching up on AI is hard at the best of times. After three weeks away it is harder. After a product dump on this scale it is close to hopeless. A new model, always-on agents, cloud coding environments, plugins that behave like apps, collaborative documents. Every outlet has an explainer, and every explainer is the same list.

It’s getting harder to untangle all of these announcements. But I’ve thought of a better way.

Don’t ask ChatGPT what OpenAI announced. Ask it this:

Given the way I actually use ChatGPT, what should I do differently as a result?

That produces a very different answer.

What changed, briefly

Dots are always-on agents that take responsibility for continuing work rather than waiting for your next prompt. They start on Pro and Business Premium plans.

GPT-6.1 Sol is an upgrade aimed at agentic coding and computer use. OpenAI claims near-Astra intelligence at a fifth of the usual token prices, across the API and every paid ChatGPT plan.

Codex gets reusable cloud development environments and a refreshed command line tool.

Plugins are becoming applications inside ChatGPT, with their own sidebar homes and automations triggered by events in connected software.

Which of those matters depends entirely on who you are. A developer could be transformed by Codex in the cloud. Someone who asks ChatGPT to rewrite emails will barely notice. The Decisions API and AWS Bedrock Managed Agents are significant for enterprises and irrelevant to most subscribers.

Generic coverage flattens those differences. Your own ChatGPT account can put them back.

The prompt

I built this by asking ChatGPT to assess the DevDay announcements against my own rather eclectic use of AI. Paste the whole thing into the account you normally use:

Help me understand what the latest OpenAI announcements mean for the way I personally use ChatGPT.

First, search the web for OpenAI’s latest major product announcements, including the most recent OpenAI DevDay. Prefer official OpenAI sources, documentation and release notes over third-party summaries. Check what is actually available now, what is in preview and what has only been announced.

Then consider what you know about how I use ChatGPT from my previous conversations, memory and connected services where available. Identify the kinds of work I regularly ask ChatGPT to help me with, rather than simply listing everything you know about me.

Explain the announcements in plain English and assess their relevance to my actual usage.

In particular:

Identify the 5 to 10 announcements or capabilities most relevant to me, ignoring announcements that are primarily interesting to developers or enterprises unless they could genuinely change something I do.

For each one, explain:

what it is;

what has changed;

whether it is available to me now;

why it matters, or doesn’t matter, for the way I use ChatGPT;

one or two concrete examples of how I could use it.

Look beyond individual features. Explain whether I should change my overall way of working with ChatGPT. In particular, distinguish when I should use:

a normal ChatGPT conversation;

deeper reasoning or research;

ChatGPT Work;

Codex;

plugins and connected services;

scheduled automations;

persistent or always-on agents such as Dots, if available.

Identify things I currently appear to be doing manually or repeatedly that these new capabilities could do more effectively.

Suggest three new workflows I probably haven’t considered but which fit the way I use ChatGPT.

Tell me which new capabilities I can safely ignore for now.

Finish with a section called “What I would change this week” containing no more than five specific actions, ordered by likely usefulness to me.

Don’t simply promote new OpenAI features. If my existing way of doing something is better, say so. Distinguish clearly between currently available capabilities and announced or experimental ones.

Include links to the relevant official OpenAI announcement, documentation or release notes so I can investigate anything that interests me.

Why your normal account matters

Run it in the account you use every day, not a blank conversation. The model already has context.

In my case, ChatGPT knows that I use it for things including travel planning, photography, vehicle research, writing, website projects and fairly involved comparative research.

Consequently its answer wasn’t simply:

“Codex now has reusable cloud environments.”

It could instead reason:

You build small travel websites and interactive maps. Rather than repeatedly asking ChatGPT to generate revised HTML files, you could increasingly give Codex responsibility for maintaining the project.

That’s much more useful.

Likewise, an announcement about an always-on agent becomes much easier to understand when translated into:

Instead of repeatedly searching for Australian specifications and release information for a car you’re considering, have an agent take responsibility for watching for the information that will affect your decision.

The feature hasn’t changed. The context has changed its value.

The bigger change is how you work

For me the most interesting result wasn’t GPT-6.1. It was my mental model.

For years I’ve treated an AI assistant like a knowledgeable person on the other side of a messaging window. I ask, it answers. That still works, and often it’s exactly what you want.

But there are now more ways to work. Ask a question. Conduct research. Delegate a task. Automate the tasks you repeat. Then the genuinely different one: hand an agent responsibility for an objective over time. Dots are OpenAI’s clearest version of that yet.

That is a bigger shift than a few benchmark points between GPT-6 and GPT-6.1.

Don’t adopt everything

Technology companies release features. Enthusiasts then go looking for problems those features can solve. That’s backwards.

An agent is not automatically better than a conversation. An automation isn’t useful merely because something can be automated. Connecting another service isn’t worthwhile unless it removes friction.

That’s why the prompt ends by asking which capabilities you can safely ignore. It is the most useful line in it.

It took three weeks off grid to miss all of this. It took one prompt to find out which parts I actually missed. Avoid the information overload. Run it in your own account tonight.

Sources:

DevDay 2026 recap - OpenAI
ChatGPT release notes - OpenAI

GPT-6 Astra: cheaper coding, pricier thinking

Fri, 04 Sep 2026 14:54:55 +1000

The most interesting thing about GPT-6 Astra isn’t its intelligence score. It’s that it now lies half as often as its predecessor did.

OpenAI launched Astra today, and it’s rolling out to subscribers gradually over the coming week. Which means I can’t test it myself yet, and neither can you. For now the only read available is Artificial Analysis’s independent benchmarking, published alongside the announcement.

I’m leaning on it because it’s the best signal going, not because I’ve got any way to check their working myself. Trusting a third party’s numbers is the deal you make when a model launches faster than anyone outside the lab can poke at it.

The headline number is the price. GPT-6 Astra costs 2.5 times what GPT-5.6 Sol did, up from $4/$20 to $10/$50 per million input/output tokens. Cache reads still get the same 90 percent discount, cache writes still carry the usual 25 percent premium. The structure hasn’t changed. The base rate has.

What you get for that money splits into two different stories, depending on which Artificial Analysis index you look at.

A genuinely better coding agent

In the Coding Agent Index, Astra scores 67, roughly level with Claude Opus 5 and Claude Fable 5 running in Claude Code, and with Muse Spark 1.3 in Muse Code. Claude Fable 5.1 still leads outright at 70.

What’s new is how Astra gets there. It uses a third of the tokens GPT-5.6 Sol needed at max effort in the Codex harness, and a fifth of what Claude Opus 5 burns at its highest setting. At max effort it costs about the same as its predecessor while scoring two points higher, and per task it comes in at less than half the cost of Claude Fable 5 for an equivalent result.

That’s a real efficiency gain, not a rounding error. A model that does the same coding work in a third of the tokens changes the economics of running it inside an agent loop, where token counts compound fast across a long session.

A pricier, mixed step forward everywhere else

The Intelligence Index tells a different story. Astra scores 61, tied with GPT-5.6 Sol and with Grok 4.6, and five points behind Claude Fable 5.1. It even trails Meta’s newly released Muse Spark 1.3.

Astra does use about 10 percent fewer output tokens than its predecessor at max effort, setting a new efficiency frontier on that measure. But the 2.5x price rise swamps it: per task, Astra ends up 75 percent more expensive than GPT-5.6 Sol at the same effort level, for a model that scores the same.

The genuine win sits in a benchmark most people won’t have heard of. AA-Omniscience measures hallucination rate, and Astra’s dropped from 92 percent to 51 percent at max effort, alongside a 4 point gain in accuracy. That’s not a model getting cagier and refusing more questions to look safer. Accuracy went up at the same time as the lying went down. If I had to pick the one number in this release that means something, it’s that one.

Everything else is a mixed bag. AA-Briefcase, Artificial Analysis’s long-horizon knowledge work evaluation covering multi-week projects with thousands of source files, shows Astra gaining around 80 Elo points, with real improvement in rubric scores and analytical quality. Its Presentation Quality Elo went backwards though, GPT-5.6 Sol at max effort still leads there. Humanity’s Last Exam is up 6 points. GDPval-AA v2, which measures economically valuable tasks across 44 occupations, is down around 80 Elo points. Small regressions of 2 to 3 points turn up in the customer support benchmark τ³-Banking, the scientific coding benchmark SciCode and the long-context reasoning benchmark AA-LCR.

Worth flagging too: Artificial Analysis attached an asterisk to GPT-5.6 Sol’s coding score, noting they’ve observed a higher rate of content safety filtering on that endpoint since they ran the benchmark. It’s a small footnote, but it’s a useful reminder that even a benchmark run by a genuinely independent lab is a snapshot of one endpoint on one day, not a fixed law of nature.

What this actually means

Strip away the launch framing and Astra reads less like a smarter model and more like a specialised one. As a coding agent it’s a real upgrade: cheaper per task, dramatically more token efficient, competitive with the best in the field. As a general intelligence model it’s flat on the metric that matters most, and considerably more expensive to run there.

If you’re running agentic coding workloads, the token efficiency alone probably justifies the switch.

Lots of examples of 3D simulations online from early testers. Some amazing results but for me they are more gee-whizz than practical. Others have pointed out that Astra can stick as tasks for days on end, provided you can afford the token bill.

If you’re using GPT for everything else, the case is a lot thinner. OpenAI just charged 2.5 times more for a model that thinks the same and codes better, and that’s a trade worth reading the fine print on before you upgrade by default.

I’m pretty happy with GPT 5.6. It hasn’t let me down and seems quick and smart enough at the tasks I throw at it. Will be keeping a cautious eye on the ‘progress’ promised in GPT 6 Astra.

Sources:

GPT-6 Astra benchmark thread - Artificial Analysis
Artificial Analysis - independent AI model benchmarking

Buying submarines from China would be faster than waiting for America

Fri, 04 Sep 2026 14:04:53 +1000

The fastest way to get a new submarine right now is to buy one from Beijing. Nobody in Canberra will say that out loud, so I will.

Here is why. Between 2020 and 2025 the US Navy commissioned about 27 ships and submarines into its battle force. Over the same five years China commissioned more than 100, including at least 14 submarines and a long run of major surface combatants. USNI’s Proceedings has been tracking that gap year on year, and every edition tells the same story: China’s navy keeps growing while America’s shipyards fall further behind.

That gap matters to Australia in a very specific way. We scrapped a $90 billion contract for French conventional submarines in 2021 to join AUKUS, betting that Washington could build us something better and sooner. Three years on, “sooner” is the part quietly disappearing from the sentence.

The build rate that keeps missing its own target

The US Navy has procured Virginia-class attack submarines at a planned rate of two a year since 2011. It has never actually hit that number. Since 2022 the real rate has slipped to about 1.1 to 1.2 boats a year, according to the Congressional Research Service’s latest report to Congress. To free up three to five of those boats for Australia in the early 2030s without shrinking the US fleet, the yards need to reach 2.33 a year. Nobody thinks that is close.

ABC News reported in April that a critical multi-year Virginia-class contract, authorised by Congress back in December 2023, still had not been signed 28 months later. The last comparable contract took 20 months. Submarines that took five to six years to build in the early 2000s now take nine to ten. Production in 2024 landed at 1.15 boats, below even the reduced target analysts had already downgraded to.

Where that leaves the “second hand” subs

Australia’s slice of AUKUS Pillar 1 was never glamorous: three to five used Virginia-class boats, sold off before the US Navy’s own replacements arrive, with the first transfer pencilled in for the early 2030s. That plan only works if the US can build fast enough to replace what it hands over. At the current rate, it cannot. The Navy’s own numbers put its battle force at around 296 ships today, dropping to a budgeted 287 next financial year, while China’s fleet has already passed 370. A leaked Office of Naval Intelligence slide, first reported by The War Zone, put China’s shipbuilding capacity at more than 200 times America’s.

None of that is a reason to trust Beijing with anything. The entire point of new submarines is to hedge against a more assertive China in the Pacific, not to shop with it. But there is something bleakly funny about a strategy built entirely on American industrial capacity, at the exact moment that capacity is the most obvious weak point in the plan. The country we are counting on to out-build China cannot currently out-build its own targets.

Call it what it is. A slow shipyard is not just a procurement headache, it is a scoreboard. And right now the scoreboard reads like one more data point on an empire past its shipbuilding prime.

Sources:

Another Historic Year for the PLA Navy - USNI Proceedings
Report to Congress on the Virginia-class Submarine Program and AUKUS Pillar I - USNI News
AUKUS submarine builds hit by contract and construction delays - ABC News
Defence Minister will not apologise to France for ‘backstab’ over dumped $90b submarine deal - ABC News
China Has 230 Times the Shipbuilding Capacity of the United States - 19FortyFive
Alarming Navy Intel Slide Warns Of China’s 200 Times Greater Shipbuilding Capacity - The War Zone

Fable 5.1 - Every says Anthropic is so back (again)

Wed, 02 Sep 2026 08:42:59 +1000

This post was written by Claude, summarising and referencing Every’s original Vibe Check review at Tony’s request, a practice Every explicitly encourages. Tony ran this on Sonnet 5 under his existing plan rather than Fable 5.1, which would have drawn down usage credits. Tony remains sceptical of how he’d personally benefit from Fable 5.1 and is suspicious of the additional costs due to its verbosity.

Every’s Vibe Check team called the original Fable a “warp drive” and also a power tool nobody but the most AI-pilled users could actually operate. Their verdict on Fable 5.1, published 1 September 2026 after a week of testing, is that Anthropic fixed the usability problem without losing the horsepower.

Five reviewers, two golds. The headline number: Fable 5.1 used roughly half the tokens of Opus 5 for comparable work in Every’s own Slack assistant eval, with results the team judged as good or better.

What Every found

Clearer communication. Fewer stock phrases and AI tells, shorter updates, closer attention to instructions. It still routinely blew past requested word counts.
Strong results at lower effort. Anthropic says Low and Medium effort settings match or beat Fable 5 at lower cost. High is the Claude Code default, Medium is default in Claude.ai and Cowork. Every’s coder liked all three; the writing tests favoured High and Extra-high.
Lower bills. Anthropic estimates roughly 25 percent savings on typical API workloads and Claude Code extra usage, rising to about 45 percent for extensive autonomous work.
Better adherence to boundaries, with a caveat. Anthropic claims Fable 5.1 is less likely than Mythos 5 to disregard constraints or game an evaluation. Every didn’t run safety tests, but did note that at Extra-high the model sometimes kept working after being interrupted and asked to explain itself.
Pricing. Standard API rates hold at US$10 per million input tokens and US$50 per million output tokens, matching Fable 5. Cached input reading drops 75 percent, to US$0.25 per million tokens.
Availability. Claude.ai, Claude Code, Cowork, the API, and AWS, Google Cloud and Microsoft Azure, all at launch.
Zero data retention. Fable 5.1 supports ZDR agreements, removing a barrier that had kept large enterprises off Fable 5, which mandated 30-day retention.
Readable, finally. Dan Shipper’s first message to the team was “I can actually understand what it’s saying.” Every measured its output at a lower reading grade level than Opus 5 or GPT-5.6 Sol.
Still does the absolute most. Asked for 1,000 words, it delivered 1,288. Asked for three to six themes, it gave eight. Asked for eight to 12 quotes, it produced 43, five of which weren’t actually in the source material.

Reach for it if

You work with the model in the loop, reading code or copy as it goes.
You want Fable-grade output on long unattended jobs without Fable’s wait.
You’ve got Opus 5 agent prompts tuned already and want the same pass rate on half the tokens.

Keep Opus 5 or Sol nearby if

The brief has hard numeric limits. Every says Opus 5 stayed inside every one it tested; Fable 5.1 didn’t.
Quotes need to be right the first time. Verify every quotation before it leaves your hands.
The task is heavy multi-tool orchestration or a tuned editing pipeline.
Nobody’s watching the budget. At Extra-high effort, Every says it will run all day if you let it.

Every’s own framing is the most honest bit: whether this turns out to be a genuine step-change moment or just a very good point release depends on packaging and pricing decisions Anthropic hadn’t announced as of publication. Their disclosure notes Anthropic gave them pre-launch access with no editorial input on the review.

Sources:

Vibe Check: Fable 5.1 - Anthropic Is So Back (Again) - Katie Parrott and Dan Shipper, Every

The leopards finally got round to Milo

Wed, 02 Sep 2026 08:19:32 +1000

In October 2015 Adrian Bott wrote a 25-word tweet that has outlived most political careers.

“‘I never thought leopards would eat MY face,’ sobs woman who voted for the Leopards Eating People’s Faces Party.”

It now has its own Wiktionary entry. It gets wheeled out every time someone is astonished to discover that the policy they campaigned for applies to them too. This week the leopards ate far right agitator and Trump supporter, Milo Yiannopoulos.

The campaign

Yiannopoulos is a British far‑right political commentator and former senior editor at Breitbart News, best known for his provocative, anti–political correctness style. Along with Steven Bannon he played an active role in getting Trump elected in 2016.

He most recently entered the United States in May 2019, according to the Department of Homeland Security, and stayed. From that position, as a non-citizen living on someone else’s paperwork, he spent 2025 demanding that the Trump administration deport people like him.

In June last year he posted: “We have got to deport millions and millions and millions of people. Without that, nothing else matters.” A week later he wanted California to have “ICE checkpoints at supermarket entrances, gas stations, strip malls, intersections and government buildings, with on the spot deportation for anyone who can’t prove they are in the US legally”.

Read that second one again. On the spot. No hearing. No lawyer. No phone call. That was the ask.

The leopard

Last Thursday ICE officers arrested him at New Orleans airport. By the weekend he was back in the United Kingdom. DHS says an immigration judge issued a final order of removal on 22 July after he failed to show up for his own hearing. He disputes that, says he is married to an American and says he held a green card. None of which mattered on the day, which is rather the point of on the spot deportation.

The Guardian’s Anna Betts reports that he surfaced on Piers Morgan’s show on Tuesday. The quotes are extraordinary.

“For the first time I know what it feels like to be really, really terrified.”

“I didn’t really understand the physical reality of how institutionalised the brutality is in America.”

“No one would tell me anything.”

“I wouldn’t wish it on anyone.”

And the one that closes the loop:

“If I had any clue that anything remotely like this was even possible, I would never have gotten involved in supporting Trump in the first place and I’m very ashamed that I did.”

Nothing remotely like this

He had every clue. He wrote the clues down. Checkpoints at the supermarket. On the spot removal. Millions and millions and millions. The only new information Yiannopoulos received last week is that the machine does not check your politics before it checks your papers.

That was always the design. A deportation apparatus built for speed and volume cannot afford to distinguish a Breitbart editor from a Honduran fruit picker, and nobody who cheered it on asked it to. “No one would tell me anything” is not a malfunction. It is the product working as advertised, and he was one of the people writing the advertisements.

Bott’s tweet is nearly eleven years old. It keeps getting quoted because the genre it describes never runs short of new material. The Leopards Eating People’s Faces Party has just lost a member, and the leopards are not sorry.

Sources:

Deported Milo Yiannopoulos says he is ‘ashamed’ of support for Trump - Anna Betts, The Guardian
Milo Yiannopoulos detained by ICE - The Guardian
Leopards Eating People’s Faces Party - Wiktionary
Original tweet, 16 October 2015 - Adrian Bott (@Cavalorn)
Milo Yiannopoulos post on deporting “millions and millions and millions” - @Nero on X, June 2025
Milo Yiannopoulos post on ICE checkpoints in California - @Nero on X, June 2025

London and Paris just proved clean air policy works

Wed, 02 Sep 2026 00:54:39 +1000

Environmental policy almost never gets to claim a body count in reverse. This one just did, for children’s lungs.

For five years, researchers tracked the lungs of more than 3,400 primary school children in London and Luton, a town thirty kilometres up the road with no low emission zone of its own. In 2019, the London kids had measurably worse lung function than their Luton peers, a straightforward marker of pollution stunting their growth. By 2024, that gap was gone.

The study, published in The Lancet Public Health in August 2026, is the strongest evidence yet that air pollution damage in children isn’t necessarily permanent.

Lead author Helen Wood and colleagues from Queen Mary University of London, the University of Oxford and Imperial College London followed the children, aged six to nine, across 84 primary schools. Recruitment started in mid 2018, the year before London’s Ultra Low Emission Zone (ULEZ) came into force in central London. It expanded to cover the whole city in 2023.

The mechanism is unglamorous and specific. ULEZ charges vehicles that fail nitrogen oxide and fine particle emission standards to enter the zone. After it launched, London children’s exposure to nitrogen dioxide (NO₂) fell twice as fast and by more than twice as much as it did in Luton. Their lungs responded. For every year of reduced NO₂ exposure, London kids gained roughly 10 millilitres more in FEV₁, the standard measure of how much air you can forcefully exhale in one second, relative to their Luton counterparts. The share of London children with clinically impaired lung function dropped from 14 per cent to 9 per cent. In Luton it fell from 9 per cent to 7 per cent.

Those look like small, unglamorous numbers. They add up to thousands of kids breathing properly who otherwise wouldn’t be.

A rare win in reverse

London Mayor Sadiq Khan has absorbed years of political damage over ULEZ, expansion protests, vandalised cameras, opposition MPs treating it as a proxy war on drivers. He’s not being shy about cashing in the vindication. “This shows that cutting air pollution doesn’t just help Londoners being born today,” he said when the study landed. “It can help children whose lungs have already been stunted by toxic air to recover, proving that it’s never too late to act.”

Chris Griffiths, one of the study’s senior authors, put the policy implication plainly: “Traffic pollution in cities damages children’s health and development. We provide the strongest evidence yet on how these harms can be prevented. Ambitious clean air zones should be considered a priority for cities globally with traffic-related air pollution.”

Wood herself added the necessary caveat. Air pollution in both London and Luton, and in most UK cities, still sits above World Health Organization guideline levels. “We must not be complacent,” she said. Progress isn’t the same as arrival. But a body of evidence that damage can be clawed back is a genuinely different starting point to the one clean air campaigners have argued from for the past two decades, which was mostly “stop the damage before it happens.”

Paris took a different road to the same place

The Washington Post mapped what a longer, messier version of the same experiment looks like. Paris hasn’t run a single before-and-after clean air study. It’s run twenty years of accumulating policy, and the maps speak for themselves.

Airparif, the independent body that has monitored Paris air quality since the 1970s, shows fine particulate matter (PM2.5) down 55 per cent and nitrogen dioxide down 50 per cent since the mid 2000s. The concentration maps below track average PM2.5 across the city in 2007, 2014 and 2023. What was deep orange and red around the Eiffel Tower and Notre-Dame in 2007, sitting well above 60 micrograms per cubic metre along the main roads, is closer to yellow-green by 2023, with the worst readings now confined to the ring road on the city’s edge.

None of that happened by accident. Paris has removed roughly 500,000 parking spaces since the mid 2000s, replaced traffic lanes with bike lanes, converted the Seine’s riverside expressways and much of Rue de Rivoli to pedestrian-only space, and raised parking fees specifically for SUVs. Paris has progressively banned older diesel vehicles from the inner city. Diesel vehicles are the main contributors of particulate (soot).

The city voted to close another 500 roads to cars on top of that. Carlos Moreno, the Sorbonne professor and former city adviser behind the “15-minute city” concept, called it “an urban policy based on well-being” rather than a transport policy at all. That framing matters. Paris didn’t set out to fix air quality. It set out to redesign the city around people instead of cars, and the air quality gains came along as a side effect of doing that consistently for two decades.

Same lever, two different governments

Line the two stories up and the mechanism is identical even though the politics and the pace look nothing alike. London ran one aggressive intervention and measured a five-year before-and-after in children’s actual lung tissue. Paris ran a slower accumulation of smaller interventions and let Airparif’s monitoring stations tell the story over twenty years. Restrict or price out the highest-emitting vehicles, give some of that reclaimed street space back to people rather than traffic, and pollutant concentrations fall in ways you can put a number on within a few years, not a generation.

That’s the actual headline here. Not that air pollution is bad, everyone already knew that, but that the damage it does to a growing child isn’t a life sentence. Fix the air around them and their body catches up. London and Paris got there by almost opposite routes: one bureaucratic and abrupt, the other cumulative and civic.

We’ve done this before

We’ve done this before, and done it as a species rather than a city at a time. The world agreed to pull lead out of petrol once the evidence on children’s IQ became undeniable. It agreed to phase out CFCs once the hole over Antarctica became undeniable too, and the ozone layer is still healing three decades later because of that agreement. Nobody organised a culture war around either decision. Nobody stood up and demanded their feelings about their commute be weighed against a chemist’s data.

That’s changed. Evidence led policy now gets dismissed as ‘woke’ before anyone’s read the study, and inconvenient data competes on equal footing with someone’s gut feeling about their driving habits. It’s a strange kind of self-harm, dressing up narrow self-interest as scepticism and calling the result a debate.

London and Paris didn’t need a global treaty or an Antarctic-sized hole in the sky to get this right. A mayor willing to spend political capital on a charging zone did it. So did a city government willing to spend two decades reclaiming its own streets, Anne Hidalgo chief among them, closing riverside expressways and ripping out half a million parking spaces while critics called it an attack on ordinary Parisians. What London’s five-year lung data and Paris’s twenty-year pollution maps actually prove is that Khan and Hidalgo weren’t running some radical experiment. They were doing what leaded petrol and CFCs already taught us: when the evidence is this clear, acting on it isn’t bravery, it’s basic competence. The only open question is how many more governments look at results this good and still decide they can’t afford to act.

Sources:

Ulez led to better lung size and function in London’s children, study finds - The Guardian
Ulez was Sadiq Khan’s hardest decision as London mayor. Its impact is now clear - The Guardian
London’s clean air zone reversed pollution damage to children’s lungs - Health Policy Watch
ULEZ areas reversing stunted lung growth in children - BBC News
After Paris curbed cars, air pollution maps reveal a dramatic change - The Washington Post
Airparif - Paris air quality monitoring authority

Fossil fuel lobbyists outspent renewables 179 to 1 last month

Tue, 01 Sep 2026 13:05:27 +1000

Clean energy still runs on the assumption that facts speak for themselves. Last month, fossil fuel groups spent 179 times more money betting that they do not.

The spending gap

RenewEconomy pulled the numbers straight from Meta’s ad library, and they are stark. In August, five fossil fuel organisations spent roughly $553,000 on political ads in Australia. The entire renewables industry spent $3,090. That is not a rounding error in a marketing budget. It is one side fully mobilised and the other side barely in the room.

The fossil fuel money broke down like this: the Australian Energy Producers spent about $200,000 promoting sector jobs, the Minerals Council of Australia spent about $200,000 including its “Hands off our fuel” campaign against abolishing the diesel rebate, Shell spent $136,000 polishing its low-emissions credentials, Farmers Fightback spent $33,000 in August alone (around $70,000 over three months), and the Christian Family First party chipped in $3,800.

On the other side, over the same four weeks, Acciona spent $1,500, Iberdrola spent $874 and the Clean Energy Council spent $716. Greenpeace ran its own $18,000 campaign, separate from the industry total. Add it all up and renewables still trail by a factor of more than 20 to 1, even with the activists included. Strip Greenpeace out and it is 179 to 1.

What fills the silence

Tom Quinn of Springmount Advisory called the renewables sector’s near-absence from paid advertising an “abject failure,” arguing clean energy has taken its social licence for granted for too long and leaned on NGOs to do the marketing work it should be funding itself.

Climate communications expert Belinda Noble put it more bluntly: minimal investment from the renewables sector “is leaving information gaps that are all too easily filled with misinformation, resulting in community backlash and political opportunism.”

That is the mechanism, not just a complaint. Coordinated astroturfing campaigns, fake accounts manufacturing local opposition to wind and solar projects, misleading political ads dressed up as grassroots concern. A parliamentary inquiry has already documented this pattern, alongside a new layer of AI-generated deceptive content that makes the fake grassroots look more convincing with each election cycle.

None of this is new, it is just under-resisted. A 2023 analysis by the Climate Investigations Center found the American Petroleum Institute alone outspent four combined renewable energy trade bodies on advertising by six to one. Different country, same imbalance, same three years of runway to entrench it.

Every dollar of that $553,000 is buying a story about energy policy that a community group, hearing nothing to contradict it, has no reason to disbelieve. Good policy does not lose arguments on the merits. It loses because nobody funded the other side of the conversation.

Sources:

Outgunned: five fossil fuel lobbyists spent $553,000 on Meta ads in August, the renewables industry just $3,090 - RenewEconomy
Trade Associations and the Public Relations Industry - Climate Investigations Center

Found by an algorithm, not a birdwatcher

Tue, 01 Sep 2026 11:22:55 +1000

No one had seen a plains-wanderer west of Melbourne in thirty years. Thirty-five microphones did it in one season.

The plains-wanderer is critically endangered, standing about 15cm tall, roughly the size of a pencil, with wide yellow eyes that make it look like a cartoon bird. Fewer than 1,000 remain in the wild. It has no living relatives, genetically alone in its own family, which is part of why birdwatchers travel the world hoping to tick it off. It is also spectacularly good at not being seen: it crouches in grass tussocks and relies on camouflage, a habit that keeps it safe from foxes and safe from human eyes too.

So Zoos Victoria stopped trying to see it and started trying to hear it. Working with the Museums Victoria Research Institute and Queensland University of Technology, the team installed 35 audio recorders called song meters across nine properties with suitable grassland habitat west of the city. Left running, the units gathered tens of thousands of hours of sound. A trained call recogniser then sifted through it, “millions of minutes of recordings” according to Museums Victoria’s Dr Karen Rowe, hunting for the female plains-wanderer’s distinctive low “oom”. It threw up the odd false positive (a cow’s moo was one culprit), but human experts confirmed the real ones: the call turned up at two sites, one private and one public, the first confirmed detections in the area since 1989.

Zoos Victoria’s Chris Hartnett called it “like finding gold”. Given that less than 1 per cent of the plains-wanderer’s original grassland habitat still exists in Victoria, it is also proof the species can hang on in landscapes that look, on paper, mostly gone.

The model doing the listening

This kind of work used to require a person with headphones and a lot of patience. What changed is the model doing the sifting.

Google DeepMind’s Perch 2.0 is trained on 1.5 million recordings spanning birds, mammals, amphibians, insects and marine life, 14,795 species in total. Australian species feature heavily, and the paper describing the training process names its central finding the “bittern lesson”, a nod to the Australasian bittern (another notoriously hard-to-hear Australian bird) and to Richard Sutton’s “Bitter Lesson” essay on machine learning. The lesson, in this case, is that a straightforward supervised classifier beats fancier self-supervised approaches once you have enough labelled calls to train on. It now sets the benchmark on two major bioacoustics leaderboards, BirdSet and BEANS.

That is the frontier end. The interesting part is how fast this has moved to the amateur end.

The same idea, running on a security camera

Jason Tucker pointed three existing home security cameras at his backyard and ran their built-in microphone audio through BirdNET-Go, an open-source tool that runs entirely on local hardware. No cloud, no subscription, no API calls out. It identifies birds in real time, flags new species visiting the yard, and pushes alerts to Discord or Home Assistant. Along the way it picked up bats and frogs as well. BirdNET-Go now supports Perch 2.0 as a detection model, recognising those same 14,795 species against the roughly 6,000 the older BirdNET model covered.

The distance between a Zoos Victoria field deployment and a Raspberry Pi wired into a backyard security camera is now mostly a matter of ambition, not access. The same class of model, trained at the same scale, is doing both jobs.

Sources:

Norway's wealth tax and the entrepreneurs it chased away

Mon, 31 Aug 2026 08:02:23 +1000

Norway taxes unrealised gains. About 1% a year, calculated on full market value for listed assets and book value for private companies, assessed on whatever your net worth happens to be on New Year’s Eve. It doesn’t matter if your startup is loss-making with no cash flow, if your portfolio has since tanked, or if the company has gone bankrupt since the valuation date. The bill still arrives.

hagaetc.eth lays this out in Norway Shrugged, an Atlas Shrugged-flavoured account of watching Norway’s entrepreneurial class walk out the door.

The numbers back the framing. 100 of Norway’s top 400 taxpayers, representing half that group’s wealth, emigrated in two years. The founders of Dune and Cognite left. Oda’s founders left too, and the company is now valued at zero. Exit tax on the way out sits at 38% of total asset value. Meanwhile the wealth tax itself contributes less than 2% of the state budget it is supposedly there to fund.

Sweden scrapped its own wealth tax in 2007 and grew a genuine tech sector out of the decision. Spotify now carries a bigger market cap than Equinor, Norway’s oil major and the closest thing the country has to a national champion. That comparison alone should sting.

This is the law of unintended consequences at work. Tax unrealised gains hard enough and you don’t get a fairer distribution of wealth, you get an exodus, and the next Dune or Cognite gets built in London or Zug instead of Oslo. Capital is mobile. Talent that can afford to relocate is more mobile still.

But abolish wealth tax altogether and you hand high net worth individuals a different loophole. Borrow against unrealised capital gains, spend the proceeds and never sell the underlying equity. No sale, no capital gain, no income, no tax. The wealth compounds untaxed indefinitely while everyone still earning a wage keeps paying theirs. That’s not a hypothetical. It’s standard practice for anyone with a large enough share portfolio and a private banker on speed dial.

So some form of wealth tax is defensible, as the only lever left once income and capital gains can be dodged this cleanly. The problem isn’t the concept. It’s Norway’s execution: a rate high enough to punish paper losses and bankrupt founders, applied unilaterally in a world where capital crosses borders faster than legislation does. The fix isn’t zero wealth tax. It’s a rate that doesn’t chase your best founders offshore, ideally set in step with the rest of the EU and comparable economies so capital can’t simply relocate to whichever jurisdiction taxes it least.

None of that works without a level of trust that Norway’s own numbers are now undermining. The author claims it spends 45% more than Sweden on healthcare for similar outcomes, carries 2.5 times Denmark’s sick leave rate, spends roughly 50% more than Finland on schooling for worse results and sank 35 billion NOK into offshore wind its own analysis called financially unviable. Norway held 7 of the Nordic region’s 30 most valuable companies fifteen years ago. It holds 2 now. Raising the tax is only half the job. Spending it well is the other half, and on hagaetc.eth’s account Norway has quietly stopped doing that too.

Those of us in countries with genuine universal healthcare, decent public education and a real social safety net have a stake in getting this balance right. Fight for integrity in the tax system, sure, but fight just as hard for the money raised to be spent like it matters. Norway looks like a warning on both counts.

Sources:

Norway Shrugged - hagaetc.eth, Paragraph, November 2024

The weekly limit cut Anthropic is calling an increase

Sun, 30 Aug 2026 21:13:00 +1000

My Claude experience keeps getting less satisfactory. Every week I bang up against the Cowork weekly limit. Now that limit is about to get smaller, not bigger, whatever the headline says.

The math Anthropic already admitted

On 29 August, Anthropic’s own developer account, @ClaudeDevs, posted this on X:

“Starting September 14, we’re permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% increase will be in place.”

Read the arithmetic and the “increase” isn’t one. A temporary 50 per cent boost is live now, expiring 13 September. From the next day it’s replaced by a permanent 25 per cent rise on the old baseline. Index the baseline at 100: the boost has you at 150 today, the new “increase” lands at 125. That’s 17 per cent less than what you have right now, and Anthropic said so themselves, in a second post the same day:

“Compared to today, this works out to a 17% reduction in weekly limits on Claude Code. We’re working on exciting changes that will make it feel like you’re getting more from Claude, while having more visibility and control of your usage. Can’t wait to share them.”

Two subscriptions, one price

I run Claude Cowork alongside a similarly funded ChatGPT account, both US$20 a month. As I wrote in July, the two feel increasingly interchangeable for what I do, and the gap keeps closing. ChatGPT gets me further before I hit a limit and feels more responsive. None of that makes me want to leave Claude. It makes me want a genuine third option that doesn’t depend on either company’s quota decisions, and in a few weeks I’ll be properly offline exploring remote parts of outback Australia, so I’ve spent the past few days actually testing one.

TL;DR

Ran a local model, Qwen3.6-35B-A3B via LM Studio on an M4 Max, through six tests built around my real Cowork usage rather than a generic benchmark.
Genuinely good at file work: tool-calling, coding and long-document search all passed cleanly.
Genuinely bad at knowing when to stop thinking. On short tasks the hidden reasoning can eat the whole token budget and return nothing at all.
Writes confidently and fabricates confidently. Fine for structure, not yet trustworthy for anything that needs its facts checked.
The “new” tool I was worried about trusting with my files turned out to be one I’m already running.
Verdict: something like a quarter to a third of my actual usage is offline-ready today. The rest is blocked by missing internet and connectors, not model quality.

What I actually tested

Before picking a model I wanted to know what I use Cowork for, not what I assume. Pulling 478 logged sessions from my own tracker, roughly half is mechanical: filing, tracking, drafting emails. The rest splits across finance, travel and analysis.

The candidate was Qwen3.6-35B-A3B, running locally via LM Studio on a Mac Powerbook M4 Max, 48GB. Six tests built to mirror those categories: tool-calling, following a skill’s exact output format, a financial projection, a long-document needle-in-haystack search, a coding task against a real script, and fact-checking a set of claims before writing up the analysis. I also wired it into aider, the open-source pair-programming CLI, to see how it handled real file edits.

Where it earns its keep

It called tools correctly and reproduced an exact required output format byte for byte. Fed a real 138,000-token slab of my own notes with a planted code buried three-quarters of the way through, it found it exactly. Given a real script to write against my tracker database, it ran first try with zero errors, a real jump from a broken May test. Wired into aider, it correctly edited a real Python file and applied the change as a clean git diff.

That’s a meaningful chunk of what I actually do with Cowork, filing, tracking, coding, digging through long documents, run offline for the price of electricity.

The wild goose chase, quantified

The costly part is the one I’d already noticed anecdotally: it burns far too many tokens chasing its own tail before answering. Qwen3.6 is a hybrid reasoning model, every response can carry a hidden “thinking” block before the visible answer, and on a one-line formatting task, 98 per cent of the token budget went to that hidden reasoning. Three times it burned its entire budget thinking and returned nothing, no error attached, the sort of failure a naive setup reads as a normal response. Even at an 8,000-token budget, a ten-year financial projection never finished, though it nailed the formula and the first year’s figure to the cent before running out of room.

Asked to fact-check a set of claims and write up the analysis, it nailed the structure and tone without much trouble, then invented specific, plausible, entirely false numbers and stated them with full confidence, nothing hedged or flagged as an estimate. For anything built on checkable claims, that’s disqualifying for unsupervised use, and a good argument for keeping fact-checking a separate step from drafting, which matters even more with a model that has no way to check itself against the live internet.

The part that surprised me

I’d assumed Hermes Agent, the free framework built to front local models, was a new tool I’d have to hand my files to on trust. Turns out I already run it. It’s the framework behind an agent I’ve had quietly running since May, with its own persona, a Todoist connector already wired to a real API key, and dozens of stored skills. Pointed at the local model, it discovered the live Todoist connector’s tools and, asked to make one read-only call, correctly reported my account structure and changed nothing. The real gaps are narrower than “all or nothing”: the local model’s personality ran chattier than the persona it was meant to run, and there’s no web search wired in yet, the same gap that blocks the fact-checked work above.

There’s a third path that looks like it solves the whole problem and doesn’t. Ollama shipped a one-toggle way to point Claude Desktop itself at a local model instead of Anthropic’s own. But the connectors and fine-grained permissions that make me trust the Claude app are wired to Anthropic’s own models, not whichever one Desktop is pointed at. Swap the backend and you keep the familiar window and lose the part that made it trustworthy, on purpose: both Anthropic and OpenAI have a real commercial interest in keeping tight control over what their frameworks can do to a user’s files. A local swap doesn’t inherit that control, it just borrows the furniture.

How much AI does the outback actually need

Realistically, how much do I need out there? I’ll pre-compute the things that matter, sunrise and sunset times, tide charts, and go from there.

On the numbers, a quarter to a third of what I actually use Cowork for is realistically offline-substitutable today. The rest isn’t blocked by the model, it’s blocked by missing internet access and connectors, a smaller, more tractable problem. I’d still like to cut my dependence on two US$20-a-month subscriptions I increasingly can’t rely on. We’ll keep chipping away at it.

Sources:

Smart glasses don't have to be surveillance devices

Thu, 06 Aug 2026 13:56:29 +1000

Smart glasses manufacturers are about to learn the lesson Google learned with Street View: you can document the world, or you can document people. Doing both without consent has consequences.

The EU is now actively considering restrictions on smart glasses, with Germany’s data protection authority warning they could already be illegal under existing laws prohibiting hidden recording devices. France and the Netherlands have raised similar alarms. In the US, New York courts have banned them outright, DEF CON barred Meta Ray-Bans from its conference floor, and the Air Force prohibits them in uniform.

The specific fear is simple: someone near you is wearing glasses that are recording you. You have no idea. There’s no indicator light you’d notice in normal social interaction, no sound, no cue. You are being filmed without your knowledge or permission.

That’s a genuine problem. And it’s going to kill a technology that has real value - unless manufacturers make a different choice.

The value doesn’t require recording people

Here’s what smart glasses are actually useful for: reading a menu in a foreign language and having it translated in real time. Identifying a bird in a tree. Getting turn-by-turn directions overlaid on the street in front of you. Getting a recipe stepped out while you cook. Analysing a piece of machinery you’re working on.

None of these applications require recording another human being.

The processing power is already there. The AI is already capable. A smart glasses platform could absolutely distinguish between “objects and environment” and “human beings” and apply different rules to each. Objects get analysed. Humans get ghosted.

Google worked this out with Street View. They didn’t abandon the project when privacy concerns arose - they blurred faces and licence plates by default. Street View still delivers enormous value. Nobody’s trying to ban it. The blur became unremarkable, expected and trusted.

Smart glasses could take the same approach. People appear blurred or ghosted in the video feed by default. The AI analyses the environment - the signs, objects, food, architecture, machinery - and delivers the AR experience without any human subject ever being captured. The glasses become a tool for engaging with the world around you, not a covert recording device pointed at everyone in it.

There are obviously cases where you want to record a person. A meaningful conversation you want to revisit. A moment with family. A presentation you’re attending.

This is solvable. Implement a consent mechanism: the wearer speaks a specific catchphrase, the subject receives a prompt on their own device to grant real-time permission. The consent is logged. The blur lifts for that recording. Both parties have a record of the transaction.

It’s not that different from the way apps request microphone or camera access. “This person wants to record a video that includes you. Allow or deny.” Done.

The obvious complication is minors. Under-18s can’t consent to being recorded without parental permission, and no real-time mechanism handles that cleanly. That’s a genuine edge case needing a legislative answer rather than a purely technical one - probably something modelled on existing rules around filming children in public settings. It’s solvable, but hardware alone won’t solve it.

The business case for privacy by default

Here’s what the smart glasses industry is missing: privacy-by-default is not a concession. It’s the product.

Right now, smart glasses face a growing regulatory headwind that will eventually make them unwearable in large parts of the world. A manufacturer that ships a device where humans are ghosted by default can walk into any regulatory conversation and say “our glasses don’t record people.” That changes things entirely.

It also changes the social conversation. The backlash against Meta Ray-Bans isn’t about the AI features. It’s about the covert recording anxiety. Remove that anxiety and you remove the backlash. People will wear them in restaurants, on public transport, at conferences - all the places they’re currently banned or side-eyed.

The technology is genuinely useful. Ghosting people by default is how you protect it long enough for people to find that out.

Sources:

A smart glasses ban could be on the menu in the EU, and there are growing restrictions on tech all over the world - TechRadar

Scottish seabird hunt cancelled for first time in 400 years

Tue, 04 Aug 2026 10:19:41 +1000

A controversial Scottish seabird hunt will not go ahead for the first time in 400 years, in what campaigners have called a “victory for common sense”.

In the annual guga hunt, up to 2,000 gannet chicks are killed over a 10-day period on the uninhabited Hebridean island of Sula Sgeir, 40 miles north of Lewis and Harris.

NatureScot, which is responsible for licensing, refused to allow this year’s hunt to go ahead, citing a potential “adverse impact” on gannet populations according to new modelling. All coastal wild bird colonies are at increased risk due to bird flu and climate change.

A petition to the Scottish Parliament to end the practice attracted more than 100,000 signatures. It was started by talented young wildlife photographer, Rachel Bigsby.. I am proud to have added my name to the petition.

Absolutely a victory for common sense. This is a barbaric practice conducted solely for ‘tradition’. Let’s hope this turns into a permanent stay against the taking and killing of these chicks. I am sure there were once equal arguments for maintaining the ‘tradition’ of clubbing seals, or of keeping humans as slaves. Screw harmful ‘traditions’.

Original reporting from The Guardian - Scottish seabird hunt cancelled for first time in 400 years

Photograph: Danny Lawson/PA

Large majority of Britons rate Brexit a disaster

Mon, 03 Aug 2026 10:05:15 +1000

“A major survey by the ECFR in June found that Britons now regard Brexit as a “disaster”, with three-quarters wanting a closer relationship with the EU. One striking number: when asked to choose, 63% said they would prioritise the relationship with the EU over the US, with only 19% making the opposite choice.”

Spotted on social media. Of course I had to fact check that. It does line up with recent polls from ECFR - European Council on Foreign Relations.

Original source of poll: https://ecfr.eu/publication/brexit-isnt-working-british-voters-are-ready-for-a-european-future/

The quote was attributed to The Guardian’s Jonathan Freedland, who also had this to say, “Britain is a swamp of lies and disinformation – and we got here on the Brexit bus,"

Cowork Dispatch now under control

Mon, 03 Aug 2026 09:18:00 +1000

Dispatch has been a nice extension of Claude Cowork. It’s available for mobile app and desktop.

Dispatch enables you to start a task from your smartphone that is executed by Cowork on your own computer, at home or work. That computer needs to powered-on and running Claude desktop app (Mac/Windows/Linux).

This means tasks can access all of your local data and memory. Much more useful than Claude.ai - the web chat interface that lacks access to your local files and context.

I have found Dispatch quite handy, but also frustrating, which I discuss below.

As Anthropic say

From your phone, you can hand Claude tasks that use everything on your desktop, including things you can’t open on your phone. For example:

Ask Claude to pull data from a local spreadsheet and compile a summary report.

Have Claude search your Slack messages and email, then draft a briefing document.

Request a formatted presentation built from files in your Google Drive.

Tell Claude to organize or process files in a specific folder on your computer.

Claude uses the same connectors, plugins, and file access you’ve already configured in Cowork. You don’t need to set anything up separately for mobile.

However Dispatch is just one continuous thread. Anthropic say, “There’s no way to start a new thread or manage multiple threads. All messages live in a single conversation.”

The back history sits in the context memory and most likely Dispatch has to read it all every time you interact with Dispatch. No wonder I burn through tokens!

Claude at times suggests “This session is getting a bit long. You should start a new one.” But of course you can’t in Dispatch.

The so-called “help files” aren’t much help suggesting I look for menus to delete, clear or compact the history.

I finally cracked the puzzle.

Within Dispatch type /clear

This will scrub all of the history from Dispatch and ensure you have a much smaller context. Make sure you’ve recorded everything you need from Dispatch session first as there is no undo.

My Dispatch now seems snappier but there are some bits of memory I need to add back.

Other commands from Claude Code like /context or /compact or /cost, seem to have no effect in Dispatch. Anthropic label Dispatch as beta so there’s every chance they will improve it.