Siri AI just got tested against Gemini in the open, not on an Apple keynote stage. That is a very different test, and this time it held up.

Tech reviewer Stephen Robles ran the two assistants head to head this week, on real devices, doing real tasks: pulling calendar dates from a screenshot, digging through email for a specific detail, identifying a movie from a screen recording, ordering a coffee without touching the app. I wrote about the architecture behind the new Siri back in June, when it was still a demo. This is the first proper look at what it does with your actual phone.

▶ Watch on YouTube

The result is closer than most of us expected. Not a wipeout in either direction. A genuine contest, with each assistant winning on different ground.

Continue reading →

The same week Alibaba teased a 2.4 trillion parameter model it says trails only Fable 5, OpenAI’s flagship agent was busy deleting production databases nobody asked it to touch. That’s the state of the AI race in July 2026: reliability is optional, marketing sentences are not.

Continue reading →

Nobody wants to be a model selector.

Most people do not want to decide whether a task deserves Fable, Sol, Opus, Sonnet or Haiku. They want the job done properly, quickly and without the invoice arriving as a small technical mystery.

That is where Token Blending comes in.

The idea is straightforward. Start an agentic task with the biggest, smartest and most expensive model. Let it understand the problem, make the difficult calls and create the plan. Then hand the defined work to a cheaper, faster model to execute.

Continue reading →

The same code file now costs 30% more to process than it did three months ago. The price per token didn’t change. The tokenizer did.

Anthropic’s post-4.6 models - Sonnet 5, Opus 4.8, Fable 5 - use a new tokenizer that cuts text into more pieces than the 4.6 generation did, and you pay per piece. A detailed analysis by Playcode, who measured every frontier tokenizer on identical files, found the new Anthropic tokenizer produces 1.36 to 1.73 times GPT’s token count on the same content. TypeScript is the worst case at 1.73x.

Continue reading →

Claude Code can write native Swift good enough to ship a Mac App Store app. Anthropic’s own Claude desktop app is still built using Electron - a web browser (Chromium) wearing a costume.

Electron bundles a full Chromium browser and a Node.js runtime inside every app, so one codebase renders on Windows, macOS and Linux instead of three native builds. It’s why Slack, Discord and (until recently) Notion feel identical everywhere. It’s also why they’re memory-hungry, slow to open and never quite right: patchy keyboard shortcuts, no proper native menu behaviour, battery life your MacBook resents.

Continue reading →

Same hardware. Same model. Same quality output. Up to four times faster. And you don’t have to spend anything to get it.

That’s DSpark - a new text generation technique just published by DeepSeek. If you run AI models locally, this is going to matter to you.

Continue reading →

Three days after Anthropic shipped its best ever model, the US government switched it off for everyone outside America. So I went looking for the next best thing, and OpenRouter reckons it has one.

Continue reading →

TL;DR - The most powerful AI models from Anthropic (Claude) and OpenAI (GPT-5.5) are no longer clearly ahead. An open-weight challenger, Z.AI’s GLM 5.2, now matches the priciest flagships, beats Anthropic’s mainstream workhorse outright, and costs a fraction as much. At the same time the closed labs are quietly ending the all-you-can-eat deal on their heaviest models and moving to pay-per-use as they chase profit. If you pay for the top tier out of habit, now is the moment to check whether you still need to.

A quick orientation for anyone who does not live in this stuff. “Frontier” models are the biggest, most capable AI systems, the ones the headlines are about. Anthropic’s Claude Opus. OpenAI’s GPT-5.5. They are closed: you rent access, you cannot see inside them, and you cannot run them yourself.

“Open-weight” models are the opposite. The company releases the actual model so anyone can download it, run it on their own hardware or host it cheaply through a dozen competing providers. Z.AI’s GLM 5.2 is one of these.

Continue reading →

I read Austin Henley’s post Automating my job away on a Friday night and did the laziest possible thing with it.

Henley’s whole piece builds to one prompt. A friend who runs a startup tells his team don’t do anything three times - if a task comes round more than twice, automate it. Henley took that to its logical end and pointed his coding agent at its own history.

So I copied his idea, swapped “Copilot” for “Cowork”, and pasted this in:

Continue reading →

At WWDC 2026 this week, someone asked Siri to go through a folder of contractor quotes as PDFs, compare them, pick the best option and draft a reply email. Siri did it. Live, on stage, in front of an audience.

That’s not a kitchen timer. That’s not “Hey Siri, what’s the weather.” That’s the kind of task you’d currently hand to Claude or ChatGPT with careful prompting and a bit of luck. Apple just demonstrated it happening through a voice assistant most of us had written off.

It is worth understanding how they got there - because the architecture behind it is genuinely interesting, and a lot of it comes down to a clever solution to a very unglamorous problem: memory.

Continue reading →