FORSMILE
Issue #11Published November 1, 2026 / Covering Oct 26-Nov 1日本語で読む

This Week in AI, 8 Items — Frontier Models Aim at Finding Vulnerabilities and Drug Development, and Mistral Releases a 1-Trillion-Parameter Model's Weights

The week in one place

This week has three threads. The first is frontier models aiming at finding and fixing vulnerabilities. Google released Gemini 4 Carbon and Mistral released the weights of Mistral Large 4, and both lead with figures for defensive use. Carbon is available only to a subset of cyber defenders, and in both cases the figures are the companies' own, with no third-party reproduction that I could confirm. The second is life sciences: OpenAI released GPT-6 Rosalind, an approval-only biology model, and Anthropic reported its drug-development work with six organizations and the acquisition of Mithril Medicine (most of the research is preclinical). The third is pricing and allowances that widen use: OpenAI released GPT-6.1 Terra at $1 input and $5 output and gave ChatGPT Pro and Business subscribers up to $500 a month in API credits, and Alphabet said Google Cloud grew 103% in the third quarter. Mistral Large 4's weights, which were on the previous issue's watching list, were released on 29 October and I could read the license text. This week, the server refused my requests for many of OpenAI's announcement pages, so some topics could not become items (see watching).

Models & Products

Google announces Gemini 4 Carbon, claiming the lead on three cyber-defense benchmarks, but only a subset of cyber defenders can use it

Google released what it calls its strongest model, but not to the general public.

From an announcement dated 28 October. Google calls Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities" and says it ranks #1 on v5.0 of the Artificial Analysis Intelligence Index at 59. The figures for each stage of defense are 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (validating by exploitation) and 71.0% on CWE-bench (patching). On availability, Google says it is making Carbon available to a subset of cyber defenders starting this week and will gradually phase in additional trusted partners through its Fairwind Program; access for developers, businesses and consumers comes "as soon as possible". It is also taking part in the US government's voluntary pre-release testing. The price is the same introductory $2 per million input tokens and $10 per million output tokens as Argon, rising to $4 and $20 after the introductory period.

So What

Carbon is not something you can necessarily get access to by applying today. It is limited to the cyber defenders Google selects. The price is an introductory one, and the announcement does not give the end date of the period. The numbers come from Google's announcement and we have not reproduced them. The announcement does not say why the rollout is narrow.

Models & Products

Mistral releases the weights of Mistral Large 4, a model with about 1T total and 49B active parameters; offering it to third parties as an API needs a separate commercial license

The weights we were waiting for are out. Running it yourself is free, but selling it as an API needs permission.

Released on 29 October. The Hugging Face model card describes a natively multimodal MoE with about 1 trillion total parameters, 49B active per token (the card says 52B including embeddings and output layers, which is where the A52B in the repository name comes from) and a 1M-token context window, taking text and images. The official variants are FP8 and NVFP4. The license is the "Mistral Ecosystem Open-Weight License", a modified MIT license. Anyone who operates a "Model as a Service" business (giving third parties inference or post-training access via API or hosted endpoints) must sign a separate commercial license with Mistral before any commercial use. Internal use, and use through Mistral's official products, are outside that condition. Mistral says it leads open-weight models in coding, with 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0 (13 points above the public preview), 48% on the Artificial Analysis Coding Agent Index (using 47% fewer tokens on average) and 51% on the Cyber Index.

So What

For running it yourself or inside your company, the terms are close to plain MIT. If you offer it to other companies as an API or inference endpoint, sign a commercial license with Mistral first. From here on, even a model described as "open weight" cannot be judged usable until you read its license. The benchmark numbers are Mistral's own, and "best" and "leading" are Mistral's claims. Whether you can run a 1-trillion-parameter model on your own hardware is a separate question.

Models & Products

OpenAI releases GPT-6.1 Terra: a lightweight model for image understanding and computer use at $1 input and $5 output

OpenAI added a fast, cheap image-capable model to the API.

The OpenAI API changelog records the release of GPT-6.1 Terra on 26 October. The model page describes it as a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing. It takes text and images as input and produces text, with a context window of 1,050,000 tokens, with the model page listing maximum input tokens of 922,000 and maximum output of 128,000, and a knowledge cutoff of 1 June 2026. Pricing per million tokens is $1 input, $0.05 cached input and $5 output, and requests with more than 272K input tokens are priced at 2x input and 1.5x output. Batch and Flex are 50% cheaper, and reasoning effort ranges from none to max. Tool calling requires the Responses API.

So What

You can try image and screen-reading workloads at a low unit price. Long inputs cost more per token, so work out the cost first for anything beyond 272K tokens. We could not open OpenAI's announcement page, so we have not confirmed availability in ChatGPT. The "everyone" wording comes from OpenAI's summary.

Models & Products

OpenAI releases GPT-6 Rosalind for biology: only organizations approved for internal research can use it, at $15 input and $75 output

It is not a model anyone can use; it is an approval-only model for life sciences.

The API changelog entry for 30 October. GPT-6 Rosalind brings "specialized biological reasoning with scientific tool use to complex life sciences research", and it is available only to organizations approved for Rosalind Research, for internal research by authorized users (a trusted-access program). In the pricing table, gpt-6-rosalind costs $15 input and $75 output per million tokens ($30 and $112.50 at long context), and gpt-rosalind-research and gpt-rosalind-discovery in the same table are $5 input and $25 output.

So What

You cannot use it with an ordinary API contract. It is aimed at research institutions and companies, and we could not confirm the approval criteria from the material at hand. Anthropic also published life-sciences work the same week (next item). But the changelog and the pricing table give no performance figures for Rosalind.

Research

Anthropic reports on its drug-development work with six organizations and acquires one-person company Mithril Medicine; most of the research is still preclinical

A progress report on putting AI into the whole drug-development process, not just one step.

From a post dated 27 October. Anthropic says it has spent several months in research partnerships with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group. The figures it gives include Enveda's manufacturing-document check (on 1,061 pages Claude caught all 11 issues human reviewers had found and flagged seven they had missed, cutting human review time by 89%, from over 25 hours to under three), BigHat testing about 1,010 variants in about 10 weeks, and one Generate design for a CAR T-cell therapy that was about 70% shorter than the standard version and let immune cells make roughly 10 times as many receptors. It also says it acquired Mithril Medicine, a one-person company developing treatments for two forms of muscular dystrophy. Its founder had Claude handle contracts, logistics and experiments across 11 external labs and, in under 18 months, reached promising results in mice.

So What

Every figure was presented by the partners and Anthropic themselves. The post also says most of the research is still preclinical and has not reached human trials. The Mithril acquisition gives no price or terms. We could confirm that Anthropic's approach of putting AI into every step of drug development and OpenAI's approval-only biology model (previous item) landed in the same week, but there is nothing here to compare the two companies' results.

Funding & Corporate

Alphabet Q3: revenue up 25% year over year, Google Cloud up 103%, first-party model APIs at about 32 billion tokens per minute

Alphabet itself put numbers on the demand for Google's AI.

This is an edited transcript of CEO Sundar Pichai's remarks on the 28 October earnings call. He said Alphabet revenue grew 25% year over year to a "half-trillion-dollar annual run rate", Google Cloud revenue grew 103%, and Search query growth was the highest in five years. On models, he said first-party model APIs now process nearly 32 billion tokens per minute (40% more than last quarter and more than four times a year ago). For Gemini 4 Argon, he said developer-platform usage in the first week ramped up three-and-a-half times faster than Gemini 3, and that teams at Chrome, Ads and Search shipped 35% more code with 30% fewer errors.

So What

This is an executive's statement on an earnings call, not the financial statements themselves. Revenue in dollars does not appear in this transcript. "Annual run rate" is a way of looking at a quarter's revenue multiplied by four, not annual revenue. Google also says it is short of cloud capacity, so strong demand and being able to use it whenever you want are separate matters.

Developer Tools

OpenAI gives ChatGPT Pro and Business subscribers up to $500 a month in API credits; unused credits do not roll over

A ChatGPT subscription now lets you try the API without a separate API bill.

From the API changelog entry for 28 October and the official guide. It applies to paid ChatGPT Pro and Business subscriptions, and requires a subscription at least seven days old, being the owner of both the ChatGPT subscription and the API organization, and linking with the same email address. Monthly credits are $100 on the $100 Pro tier, $200 on the $200 tier and $500 on the $500 tier; for Business, $20 per paid Standard seat and $100 per paid Premium seat, with Business capped at $500 per workspace per month (pooled across seats). Unused credits expire at the end of the billing cycle and do not roll over. You can use the API without registering an API payment method while you have active credits.

So What

Individual developers with ChatGPT Pro can try the API at no extra cost. The guide says "while these benefits are available", so the benefit is not promised for any length of time. The credits are subject to API rate limits and terms. The free tier and Plus are not listed as eligible.

Developer Tools

Claude Code 2.1.311 to 2.1.316: fast mode for Opus 4.8 ends on 30 November, and a setting to block writes outside the working directory arrives

If you use fast mode on Opus 4.8, you need to migrate before the deadline.

The releases listed on GitHub are 2.1.311, 312, 314 and 316 (313 and 315 are not there). 2.1.316 says fast mode for Opus 4.8 is deprecated as of 30 October and will be removed on 30 November. After removal, a fast-mode request on Opus 4.8 does not fall back to standard Opus 4.8; it returns an error. It advises migrating to Opus 5.5 to keep the speedup. 2.1.314 adds `permissions.blockWritesOutsideWorkingDirectories`, so Write, Edit and NotebookEdit refuse paths outside the working directories in every permission mode. VS Code and the Claude Desktop Code tab gained idle compaction, which summarizes a conversation of 200K+ tokens left idle. 2.1.316 also adds `CLAUDE_CODE_DISABLE_ULTRATHINK`, `CLAUDE_CODE_SUBAGENT_EFFORT` and pooled group budgets for the gateway.

So What

If you use fast mode on Opus 4.8, switch to Opus 5.5 before 30 November. From that day the requests will return errors. The setting that blocks writes outside the working directory is worth turning on before you hand over automated runs. We did not run each change in these release notes ourselves.

Watching (no confirmed primary source yet)

  • Google DeepMind announced "Genie 4", a real-time world model, on 27 October. The remarks on the Alphabet earnings call describe it as simulating "safe, realistic, interactive worlds with native audio in real time" and as "an important building block for training robots and agents". But I could not retrieve the body of the DeepMind page and so could not confirm the details, and it is not an item.
  • On 29 and 30 October OpenAI published posts on detecting frontier-model misalignment, an updated Model Spec, a community safety report with new protections for teens, and a robots.txt signal for AI-training preferences. I checked the summaries in the official feed, but the server refused my requests for the announcement pages, so I could not open the full text and these are not items. The expanded Oracle and ServiceNow partnerships of 27 and 28 October are skipped for the same reason.
  • For Gemini 4 Carbon, availability for developers, businesses and consumers is only "as soon as possible". On the earnings call the CEO said new models, including more efficient versions, are coming "soon". I will check price and scope when they appear.
  • I could not open OpenAI's announcement page and so have not confirmed which ChatGPT plans include GPT-6.1 Terra. I will cover it when I can.
  • The previous issue's item on OpenAI models and Codex coming to Google Cloud's Model Garden (general availability "in the coming weeks") had no follow-up in this week's official feed.
  • I checked Anthropic's expansion of AI for Science (28 October; Claude credits for 20 foundations and research institutions) and NVIDIA's new RTX Spark laptops and Nemotron 3.5 Transcribe (28 October), but there is too little to say what changes for the reader, so I skipped them.
  • On funding and M&A, the only deal confirmed by the party itself was Anthropic's acquisition of Mithril Medicine, and no price is given.

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues