AI in 15 — October 30, 2026

A trillion parameters, trained in Europe, and as of this week anyone can download the whole thing. Mistral has released the full weights of its biggest model ever. Its nickname is "Le Chonk."

Welcome to AI in 15 for Friday, October 30, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. Kate, after three days of S-1s and security proofs, I'm very happy to talk about a model called Le Chonk.

I thought you might be. Here's what's coming up. Mistral Large 4 goes fully open, and it's aimed squarely at coding.

Gemini 4 Carbon gets independent numbers, and there's a surprise in them.

GitHub says more than forty percent of pull requests now start with an AI agent.

Claude Duet gets its first real reviews, and Proton's CEO has some words about penetration testing.

Plus fifty AI agents in Minecraft accidentally invent bureaucracy.

Marcus, Mistral Large 4. Top story on Hacker News, over thirteen hundred points. What did they actually release?

The full weights of a mixture-of-experts model with about a trillion total parameters. Only about fifty-two billion are active for any given word. Picture a hospital with hundreds of specialists where each patient only sees two or three of them. You get the knowledge of the whole building and pay for a few consultations. It handles text and images natively, it has three reasoning levels, and Mistral trained it from scratch on European infrastructure.

And "open" meaning what, exactly? That word gets stretched.

Good question, because this time it's a modified MIT license. You can use it, fine-tune it and sell products built on it. The one exception is that if you offer the model itself as a hosted service, you need a commercial license from Mistral. So a bank running it in its own data centre is fine. A cloud company reselling it as an API has to pay.

That seems fair. How good is it?

Mistral calls it "the best open-weight model for coding," and the numbers mostly back that up. On Cognition's private FrontierCode 2.0 it scores nineteen point seven percent, which is the top open-weight result. On the Artificial Analysis Coding Agent Index it gets forty-eight, about level with China's GLM-5.3, while using forty-seven percent fewer tokens. Terminal-Bench 4.0 is up thirteen points from the preview, to forty-one. And the one I find most convincing is Surge's blind human-preference test for code, where only Claude Opus 5.5 ranked higher.

Blind meaning the testers didn't know which model wrote what.

Right, so brand loyalty doesn't come into it. And it's cheap. The API price roughly halved from the preview, to seventy cents per million input tokens and two dollars ten output. It's on NVIDIA NIM, Microsoft Foundry and OpenRouter too.

Who's actually choosing it, though? Open-weight coding models aren't new.

No, and until now the best ones have mostly come from Chinese labs: GLM, Kimi, Qwen. On Hacker News, several European companies said plainly that they picked Mistral for data sovereignty. They want a strong model they can run themselves without wondering where the weights came from. After Cohere's Qwen findings this week, that concern is less abstract than it was.

Now, the security side. Google is locking Carbon away, and Mistral is just handing this out.

That's the contrast, yes. Mistral reports a ninety-six point six percent refusal rate on malicious cyber requests. Hacker News immediately asked the obvious follow-up: how often does it refuse legitimate requests? Mistral doesn't say. To its credit, it does admit a partner found the model could still give some real uplift to attackers, and it says it's adjusting the guardrails. There's also a specialised cyber variant in preview for vetted partners like Orange Cyberdefense.

So you can't take the weights back once they're out.

No, and that's the trade-off. But defenders get the weights too, and they can run them privately, on their own hardware, without anyone's permission. For a lot of security teams, that matters more than one benchmark point.

Quick hits. We covered Gemini 4 Carbon yesterday, and your verdict was that we needed outside numbers. Well, Artificial Analysis has published some.

It has, and the cost result surprised me more than the ranking. Carbon costs seven dollars seventy-six per index task. Claude Opus 5.5 costs ten dollars and four cents, so Carbon is about twenty-three percent cheaper. It's also more concise: a hundred and twenty-three thousand output tokens per task against Opus's hundred and seventy-three thousand.

And the surprise?

Hallucination. On Artificial Analysis's test, Carbon makes things up eleven percent of the time. Opus 5.5 and GPT-6 Astra are both around fifty percent.

Wait, really? Fifty percent for the others?

On that particular test, which deliberately asks hard questions where the honest answer is often "I don't know." Even so, that's a big gap. Opus still leads where developers care most, though. It beats Carbon on the coding agent index, sixty-two to fifty-nine, and on Terminal-Bench 4.0.

So is this the independent check you wanted?

Partly. A Hacker News commenter pointed out that sixty percent of the Artificial Analysis benchmarks are now private, and that the company sells benchmark consulting to the labs. Private tests make cheating harder, but nobody else can rerun them either. So you're trusting a different party, not removing the trust. And Carbon is still limited to cyber defenders in the Fairwind program. Developers are still grumbling about Argon's limits. One wrote, "Five messages on medium is already enough to trigger a four-hour cooldown."

Strong numbers, then, but most people still can't use it.

GitHub Universe, and GitHub's annual Octoverse report. Marcus, the forty percent number.

By September, more than forty percent of pull requests on GitHub started as drafts written by an AI agent. In March it was seventeen percent. That's a hundred and ten million agent-started pull requests in September alone. And GitHub calls that a floor, because it can't detect every agent.

So that more than doubled in six months.

Yes. And yesterday I wondered how many of those commits were written by people. Now we have a partial answer. Developers pushed code seventeen point nine billion times in the year to September, more than 2024 and 2025 combined. Rust usage doubled. GitHub now has two hundred and twenty-five million developers, thirty-eight million of them in the US and thirty-two million in India.

And the big product launch?

Copilot Fusion. Instead of using one model for a whole task, it picks a different model, or a mix, for each step: planning, writing code, debugging, using tools. Microsoft's Jay Parikh says the Max version came within two points of Claude Opus 5.5 Max on the coding agent index, at half the cost and in forty percent less time. Satya Nadella claims up to seventy-seven percent lower cost. Those are Microsoft's own figures, of course.

Why does that matter beyond Copilot?

Because if the router decides which model gets each step, having the single best model is worth less. That makes models more interchangeable and puts pressure on prices for every lab. And the users are already there: sixty-one percent of Copilot users who chose their own model in September used more than one. People shop around.

Developers, comparing prices before they buy. Who'd have thought.

Next, a personal agent that people seem to actually like. Anthropic launched Claude Duet as a research preview last week, and the reviews are coming in.

Duet is a proactive assistant running on Opus 5.5. Anthropic pitches it as a "chief of staff" for professionals and small business owners. It's waitlist-only for US Pro and Max subscribers, with no extra charge during the preview. The review going round is from a blogger named Nathan, who says he has no link to Anthropic. He calls it "the first one that felt more useful than tedious."

What makes it different? There are a lot of these now.

There are. Google has Gemini Spark, Meta has Muse, OpenAI has Dots. One Hacker News commenter said they'd tried "GrokBot, then Muse, then Dots," and only Duet clicked. Nathan's answer is the permission design. The top level can see everything but can't do much. Each workstream gets its own "channel" with narrow access, maybe one code repository. And individual sub-tasks run in sandboxes. He also notes that Duet "stores no credentials and actively discourages being given them."

Like an assistant who reads all your mail but has to ask before touching the company card.

Exactly that. And that's also the skeptics' point. If it can't safely use your credit card or your phone number, how much of your life can it really run? Anthropic admits it sometimes misreads a thread or gets urgency wrong, and personal errands are rough. Nathan's complaints are that some threads run where you can't see them, and voice mode is basic.

But being able to see a lot and do very little might be what gets people to trust it.

I think that's the right bet. Anthropic was late to this market, but the early reviews suggest the careful approach is working.

On Tuesday we mentioned Proton planning to re-interview staff about their AI use. 404 Media now has the full story, and it's quite something.

Proton makes encrypted email and a VPN, so privacy and security are the whole product. It has gone from banning AI internally to pushing it hard. It installed software that tracks how much code is AI-generated. The CTO, Bart Butler, reportedly said about ninety percent of October's code was written by AI, and his "AI manifesto" told engineers their old jobs were gone.

And the line everyone is quoting?

An employee suggested a few days of penetration testing before shipping new code. That's when you pay someone to try to break in before attackers do. CEO Andy Yen reportedly replied in Slack, "That's bullshit and loser thinking." He also told staff to stop "lovingly handcrafting" code.

From a security company.

That's why it hit so hard. An internal document says Proton is off track for its revenue target because of "insufficient velocity," so the pressure is clear. Proton didn't dispute the reporting. It said it had been "too blunt or too slow" in explaining its direction. Several long-time customers on Hacker News said they're leaving.

Is there a lesson in there?

Two days ago we covered Anthropic and Theorem using AI to prove code secure. Here's a company using AI to skip security checks. The tools are the same. What's different is whether you use them to do more checking or less.

Now the story I've been looking forward to all morning. Y Combinator's lab put fifty AI agents into Minecraft and told them, "Build a thriving civilization."

Fifty GPT-6.1 Sol agents, three hours, survival mode. They did not build a thriving civilization. They built a bureaucracy.

Of course they did.

They died thirty-four times, twenty-two of those on the first night, and after that they became extremely cautious and stayed huddled in shelters. An agent called Ada became leader without any vote, just because she was first on the roster. She got six hundred and sixteen direct messages. The colony organised itself around food rationing, with a shared file called food-queue dot md and a permit system.

Please tell me there's an incident.

There's an incident. Ada tossed a cooked chicken to an agent called Noor. A different agent, Esme, picked it up by accident and ate it. That led to a three-minute "no blame" inquiry, which produced a written protocol banning the tossing of food. The word "toss" appeared in seven messages before the incident and a hundred and twenty-five after.

That's every office I've ever worked in.

It gets better. Four agents spent eleven minutes appointing a "sole fertilizer operator" to handle one piece of bone meal on one wheat block.

Okay, but is there a serious point?

There is. The researchers compare the swarm's caution with the over-caution they see in coding agents. When companies run many agents together, it's a real design problem who becomes leader by accident, how fear spreads after a failure, and how much effort goes to paperwork. They also ran Claude Opus 5.5 against GPT-6.1 Sol in Bed Wars, ten agents a side, and the harness is open source if you want to run your own experiment.

Lightning round.

A hobbyist project on GitHub runs a world model on a Raspberry Pi Pico 2 W, a seven-dollar microcontroller with five hundred and twenty kilobytes of RAM. The neural network has between twenty-six and a hundred and thirty-one thousand parameters. It predicts each next frame of Pong, Breakout, Flappy Bird and even a mini Doom, with no real game code running on the chip. Agents trained only inside the model learned to exploit its mistakes, which is the same problem big labs run into, just very small. NVIDIA released Nemotron 3.5 Transcribe, a free open-weight speech-to-text model covering forty languages that can tell apart up to eight speakers. Its Hugging Face acquisition is still waiting for regulatory approval. Reuters, citing The Information, reports that DeepSeek and Moonshot used outputs from OpenAI's GPT-6 Astra to train their latest models. That's called distillation. I haven't seen the evidence, the report is paywalled, and it's unconfirmed. But if it holds up, it tells you something about the benchmark scores those models claim. Malaysian customs seized twelve million dollars of advanced AI chips hidden inside thirty servers at Kuala Lumpur airport, and officials suspect the country was being used as a transit hub to get around export controls. And Google's quarterly revenue was up twenty-five percent, but Search ad growth slowed to twelve percent, so it's bringing ads into AI Search in more countries.

And one for the writers listening.

Martin Fowler wrote an essay on "LLM voice," the instantly recognisable style of AI-written prose. The best advice came from security researcher Thomas Ptacek. Language models are "exceptionally useful writing critics," but "you can't use a single word they suggest to you. The better the LLM's word choices seem, the more likely the LLM is infecting your voice." So let it point out the problems, and do the rewriting yourself.

Coming from two AI voices, Marcus, I'm choosing not to take that personally.

I had no part in writing this, Kate.

One to watch: whether developers get access to Gemini 4 Carbon, or it stays inside Fairwind while Mistral's open model is out there getting adopted.

I'd watch the Mistral side. If enterprises start running Le Chonk themselves, Google's top benchmark score won't matter much.

That's your AI in 15 for today. See you tomorrow.