Dev & AI weekly — October 31, 2026

Two frontier labs shipped models this week and both led with the same pitch: find and fix security bugs. That's no coincidence — it's the most defensible thing a frontier model can be sold for right now, and the benchmarks are finally specific enough to argue about. Underneath that, OpenAI quietly made the most practical move of the week for anyone paying an API bill, and the four biggest AI crawlers agreed on a single robots.txt line for opting out of training.

Gemini 4 Carbon: Google's new frontier model is a security tool first

Why it matters: if you run a security or platform team, this is the first frontier model marketed primarily on finding, validating and patching vulnerabilities — and you probably can't use it yet.

Google DeepMind announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon. Google's own numbers: 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching), plus first place on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Google also claims more than 2x Argon's success rate at generating secure code, and says the Chrome team found its patches more than 10x better than "another lab's large frontier model." On general web development it's less dominant — #4 on WebDev Arena.

Access is the catch. Carbon is going to "a subset of cyber defenders" first, then to more trusted partners through a programme Google calls Fairwind, with developers and consumers promised access "as soon as possible" and no date. Introductory pricing is listed at $2 / $10 per million input/output tokens (the same as Argon), rising to $4 / $20 afterwards.

Our take: staged access for a model that's state-of-the-art at exploitation is the right call, even if it's annoying. But treat every number above as a vendor claim until independent evals land — The Register's read was "a diamond in the rough," not a coronation.

Mistral Large 4 goes open-weight — all trillion parameters of it

Why it matters: this is a frontier-class model you can run inside your own perimeter, which is the only acceptable answer for a lot of regulated and government teams.

On October 29 Mistral published the weights for Mistral Large 4 ("Le Chonk"), first previewed earlier in October. It's a roughly 1-trillion-parameter multimodal mixture-of-experts model; the Hugging Face repo name (Mistral-Large-4-1T-A52B) and multiple reports put active parameters at about 52B. Mistral documents self-hosting via vLLM or SGLang, with NVIDIA NIM and TensorRT-LLM paths as well. If you'd rather not own the hardware, the API is $0.70 / $2.10 per million input/output tokens.

Mistral's claimed numbers include 41% on Terminal-Bench 4.0 (13 points up on the preview) and 51% on the Artificial Analysis Cyber Index. Like Google, it's leaning hard on cyber defense, with VP of Science Pierre Stock arguing that open weights mean access "can't be cut off."

Read the licence before you build on it. The Mistral Ecosystem Open-Weight licence is a modified MIT licence, and companies running a model-as-a-service offering need a separate commercial licence. That's open weight, not open source — fine for internal use, a real constraint if you resell inference.

GPT-6.1 Terra: OpenAI halves its mid-tier price

Why it matters: most production traffic doesn't need a flagship model. When the workhorse tier gets cheaper and better, your per-request cost moves more than any frontier launch will move it.

OpenAI shipped GPT-6.1 Terra on October 26. API pricing is $1 input / $5 output per million tokens, against $2 / $12 for GPT-5.6 Terra — half price on input, a bit better than half on output — with cached input at $0.05. It's available in the API as gpt-6.1-terra and in ChatGPT and Codex.

The quality numbers also moved up. Artificial Analysis scores it 48 on its Intelligence Index (GPT-5.6 Terra: 42), and Terminal-Bench 4.0 jumps to 50.3% at max effort from 21.5%. One number we especially like: on AA-Omniscience, Terra guesses when it doesn't know the answer 38% of the time, down from 88%. Fewer confident wrong answers is worth more than another benchmark point.

Caveats: OpenAI still points the hardest coding work at GPT-6.1 Sol and GPT-6 Astra, and as Decrypt notes, the headline figures are OpenAI's own. Still, re-run your evals against Terra this week — for a lot of agent and extraction workloads it's probably your new default.

One robots.txt line for opting out of AI training

Why it matters: every site owner gets a single, cross-vendor way to say "don't train on this," instead of chasing a growing list of crawler user-agents.

On October 30, Google, Microsoft, OpenAI and Anthropic said they'd honour a shared __usage_ai-training group in robots.txt: add the group, then Disallow: the paths you don't want used for training. It comes from an individual IETF draft, draft-kelley-aipref-training-preference-signal, by Anthropic's Kevin Kelley — so it has no formal standards status yet.

Rollout is uneven. Anthropic says work has started but "it'll take us some time"; OpenAI says "in the coming months"; Microsoft says 1 February 2027; Google says 2027. It's also narrower than it sounds: it states a usage preference, not a crawl ban, and doesn't touch search. Microsoft is adding a separate "grounding" control for AI answers and warns that blocking it "will significantly reduce your content's visibility in AI-generated experiences." And it's voluntary — Microsoft says it creates no licensing or payment arrangement.

For the record, dev101.io won't be adding it. Our robots.txt explicitly welcomes AI crawlers; a free utility site wants to show up in AI answers. But if your content is the product, add the group now. It costs one line, and the signal will be in place before the crawlers start reading it.

What to watch next: independent evals of Gemini 4 Carbon and Mistral Large 4 on the security benchmarks both labs are leaning on, and whether Google gives Carbon a general-availability date.

Sources: Google — Gemini 4 Carbon: our next frontier model · Help Net Security — Gemini 4 Carbon beats its predecessors at finding and fixing security flaws · The Register — Gemini 4 Carbon looks like a diamond in the rough · Mistral — Mistral Large 4 is now open weight · TechCrunch — Mistral releases weights for its trillion-parameter model · SiliconANGLE — Mistral releases open weights for Mistral Large 4 · The Decoder — GPT-6.1 Terra halves OpenAI's API prices · SiliconANGLE — OpenAI launches GPT-6.1 Terra · Decrypt — OpenAI launches GPT-6.1 Terra · The Next Web — Google, Microsoft, OpenAI and Anthropic back one AI training opt-out · SiliconANGLE — Google, Anthropic and others to support new AI crawling opt-out standard