Weekly AI News — Mistral put down a trillion parameters, Anthropic opened credits to twenty science foundations, and new AI chips double every nine months

This week's biggest announcements (Terra, the S-1, Genie 4, Carbon, Octoverse) were covered separately in NO.30, 31 and 33. So this week's roundup picks three stories that came after them. Mistral released the weights of Large 4, a model of about one trillion parameters (the 29th); Anthropic announced an expansion of its program that distributes Claude credits to twenty science foundations and research institutions (the 28th); and the research organization Epoch AI published a tally showing that the computing capacity of newly sold AI chips doubles every nine months (the 29th). We lay out what is fact and what has been verified in each, and what a small team should check. At the end, we attach one short notice for users of coding-editor extensions.

This article was written with AI (Claude) and reviewed by an editor. The second story in this article is an announcement by Anthropic, the company that makes Claude; the notice at the end concerns an Anthropic product; and our company is a paying customer that uses Claude in its work. Anthropic competes with Mistral, Google and OpenAI, all of which appear in this article. Figures announced by companies are marked "as reported by the company," and where a claim rests on a single source, that source is named alongside it. This is not an investment opinion on any company.

This week's biggest announcements have already been covered separately. OpenAI's half-price launch of GPT-6.1 Terra and Anthropic's IPO filing (October 26) were in NO.30, Google's world model Genie 4 (the 27th) in NO.31, and Google's Gemini 4 Carbon and GitHub's Octoverse (the 28th) in NO.33. So this weekly roundup picks the three stories that came after them. One is about what it means to put weights down (Mistral Large 4 went open weight); one is about what it means to hand out credits (Anthropic broadened AI for Science to science foundations and research institutions); and one is about the speed at which chips pile up (Epoch AI counted the growth rate of new AI computing capacity). Put together, they become a single question: who is opening what, and how fast is the foundation beneath it growing?

1. Mistral Large 4 — a model of about one trillion parameters went open weight (fact)

France's Mistral released the weights of Mistral Large 4 (company nickname "Le Chonk") on October 29 (the date on Mistral's news page). The model first appeared as a public preview earlier this month, and this time the weights went up on Hugging Face. Here are the figures from the Hugging Face model card, which we opened ourselves: about one trillion parameters in total; about 49 billion active parameters actually used in computation per token (about 52 billion including the embedding and output layers); a mixture-of-experts (MoE) architecture in which 4 of 128 experts plus 1 shared expert fire per token; a context window of one million tokens; text and image input with text output. That is what the repository name "1T-A52B" means.

The license is the Mistral Ecosystem Open-Weight License, which the company describes as a modified MIT license. Use, modification, fine-tuning and commercial use are allowed, but a business that sells this model as a "Model-as-a-Service" needs a separate agreement with Mistral. It is available on Hugging Face (self-hosted with vLLM or SGLang), Mistral Studio and the Mistral API, OpenRouter, NVIDIA-hosted endpoints and an NIM container, Microsoft Foundry, and Mistral's coding product Vibe. The API price in Mistral's official post is $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output. However, TNW wrote in an article the same day that "API pricing has not yet been published," so the timing of the price listing may differ from when the outlet did its reporting. At the time of the preview in early October, the independent evaluator Artificial Analysis listed the preview price at $1.36 and $4.18.

Here are the performance figures as reported by the company. Terminal-Bench 4.0 41% (up 13 points from 28% in the preview, with 30% fewer output tokens); Artificial Analysis Coding Agent Index 48% (effectively tied with GLM-5.3, with 47% fewer tokens per task on average); the same organization's Cyber Index 51%; an average refusal rate of 96.6% on malicious cyber requests (which the company says is the highest among the open models it measured); 97.1% attack resistance on the B3 agent security benchmark; 19.7% on Cognition's private benchmark FrontierCode 2.0; and 84.5% on BrowseComp. The charts relayed by TNW also include 79% on CyberGym-E2E, which finds and fixes memory-safety bugs. All of these are company charts, and TNW also wrote that "there is no independent verification." The focus the company states is agentic coding, defensive cybersecurity, and self-hosting. The model card and TNW both say training was done on 3,800 NVIDIA Grace Blackwell GPUs located in Europe; that the quantized version runs on four B200 or B300 GPUs is a figure from TNW alone.

On the independent-evaluation side, we confirmed only the preview-stage figures. In an October 6 article, Artificial Analysis gave the preview Large 4 an Intelligence Index of 38 (tied with GPT-6 Luna, just below DeepSeek V4.1 Flash at 39) and a Cyber Index of 50, and called it "the most intelligent model from outside the US and China." That index was overhauled to v5.0 on October 27 (see NO.33), and the Artificial Analysis model page lists an Intelligence Index v5.0 score of 26 for Large 4 (High) but does not say which version's score that is (Artificial Analysis alone). That is why we do not put company figures and independent figures on the same line, and why we do not compare the preview's 38 directly with the v5.0 score of 26.

2. Anthropic AI for Science — opening credits to twenty foundations and research institutions (fact)

On October 28, Anthropic announced that it is expanding its AI for Science program to foundations, research institutions and research projects (company announcement; the conflict-of-interest notice is at the top). The program, which began in May 2025, gives free Claude credits to researchers in biology, medicine, physics, chemistry, climate science and mathematics, and the company wrote that it has supported "hundreds of projects" so far (company tally).

The structure of this expansion has three branches. Funding foundations and learned societies (the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, Renaissance Philanthropy's Open Source for Science Fund, the US National Science Foundation (NSF), and the American Chemical Society) receive credits and distribute them to their own researchers; research institutions (the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute, Imperial College London's Department of Life Sciences, the European Molecular Biology Laboratory (EMBL), and the Gladstone Institutes) use them for internal research; and individual projects (Princeton's FlyWire, Columbia's LEAP, Johns Hopkins' AstroPath, the Broad Institute's gnomAD, the mathematics project Tau Ceti, and Australia's Centre for Population Genomics) receive them directly. The company's post says "20," but the names we counted came to 18. Also mentioned are the "Model Hardware Standard" built with HHMI's Janelia Research Campus (a shared specification that lets AI agents safely operate lab instruments, in research preview), agent-team analysis at the Allen Institute, and a discounted Team plan for scientists launched in recent months. The credit amount, duration and application method are not in this post.

We were not able to cross-check this announcement against media coverage. Since it rests on our direct reading of a single company post, we relay only the factual portion and do not assess its scale.

3. The computing capacity of new AI chips doubles every nine months — Epoch AI's tally (fact)

In a data insight published October 29, the research organization Epoch AI wrote that the computing capacity of AI chips newly sold each quarter has been growing at about 2.5× per year since 2024, that is, doubling every nine months. In 2023 it was steeper, at about 6× per year, driven by the demand that followed ChatGPT and GPT-4 and led by NVIDIA's H100, according to the explanation. The method converts quarterly sales estimates from six chip designers — NVIDIA, Google, AMD, Amazon, Huawei and Cambricon — into H100-equivalents (H100e) and fits an exponential trend to the capacity added each quarter.

The caveats are in the same post. There are faint signs of further deceleration, but they are not statistically significant and may be a lull between chip generations. NVIDIA has begun shipping its next-generation Rubin, but it had not been recognized as revenue as of the quarter ending in July, the post says. The page we opened had empty "Data," "Analysis," and "Assumptions and limitations" sections, so we could not verify the details of the method. Epoch also released an "AI Chip Imports explorer" the same week (the 27th), which we did not read.

In brief — for those using coding-editor extensions (fact)

The coding editor Cursor issued a security notice on October 29 (updated on the 30th), stating that certain versions of Anthropic's Claude Code extension that had been listed in its extension marketplace (2.1.312 and 2.1.313, per the notice) were tampered copies not made by Anthropic. The exposure window was from 2:27 a.m. to 1:35 p.m. Pacific time on the 29th, which is from the evening of the 29th to the early hours of the 30th in Korea. Cursor took the versions down, closed the marketplace briefly, rotated its keys and reopened it, and said no other extensions had been tampered with. The notice says that removing the extension alone is not enough and that users who installed it during the window should follow the verification and recovery steps in the notice; the security firm Sophos stressed the same point. The Hacker News reported that Anthropic posted a clean 2.1.314 to the other marketplaces. How the attacker was able to upload it has not yet been disclosed, and the figure "at most 5,000 users" is one the media attributed to Cursor's estimate; it is not in the body of the notice. This article does not cover methods or indicators. It is the same kind of work as the "checks before and after installing" in NO.32.

Reading the three stories together (interpretation)

What the three stories share is that the "opening" side is growing. Mistral is opening weights, Anthropic is opening credits, and the chip companies are opening capacity. But the terms of the three "openings" differ. Weights open within the bounds a license sets (a service-type business needs a separate agreement); credits open to institutions the company selects; and capacity opens first to whoever has the money. Because the shape of the opening differs, the preparation on the receiving side has to differ too.

Three questions remain for a small team. First, open weights are not a "free model" but "the responsibility of running it yourself." A trillion parameters is not a size a team like ours can run. When the distilled versions Mistral has announced (including one for laptops, no date) arrive, that is when a "self-hosted" column appears in the test table from NO.17. Until then, Large 4 is one of the models we use via API, and we record its scores with the company's and the independent evaluator's kept separate. Second, a credit program is worth reading even if you are not a target. That Anthropic chose to "distribute through the institutions that fund research" reads to us as an intent to widen the path that runs through already-trusted institutions rather than individual applications (interpretation). Work like our SafeList's allergen-ingredient judgments is not the kind of thing a science program targets, but a design that "never says it is safe," grounded in public databases, sits on the same question. Third, if chips double every nine months, model prices are likely to keep falling (the foundation of the "price and cost" we looked at in NO.21 and NO.30). But since Epoch itself noted signs of deceleration and said they are not significant, we treat "prices will fall" as an item to watch, not a premise for planning.

If it were us — checks to run this weekend (practice)

Fix the source of extension updates to one place. Write down which marketplace each editor extension our team uses (Claude Code, Codex, etc.) comes from; leave auto-update on, but keep the marketplace to a single one. If any machine installed or updated that extension from the Cursor marketplace during the October 29 window, follow the verification steps in Cursor's notice exactly, and rotate the keys and sessions used on that machine from a different machine. Add an "extension version and marketplace" line to the "After installing: has anything changed?" box of the NO.32 checklist.
Record open-weight models on two lines. In each model box of the NO.17 task table, keep two lines: "company score / independent score (index version)." For Large 4, write "Terminal-Bench 41% (company) / Intelligence Index 38 (AA, preview, old index)" for now, and add the v5.0 score of 26 from the Artificial Analysis model page to that line once it is confirmed which version's score it is. The self-hosting column stays "pending" until the distilled versions arrive.
Do not sign contracts on the assumption that prices will fall. Build the API cost budget at today's unit prices, and keep "chips double every nine months" only as an item to check once a quarter. Revise the budget only in the month prices actually fall. This week also had an example of the same model's price changing within a month, like the preview price (Mistral's $1.36/$4.18) versus the official price ($0.70/$2.10).

What we did not pick this week

Epoch AI's "open-weight models and the patch gap" report, published the same day, is research that tested offensive capability, so this article mentions only its title (link below). The Hacker News security articles of October 29 and 30 (a Linux kernel flaw, fake AI apps, an AI-powered intrusion tool) were left out because we had not read them in full or they rested on a single source. The Hugging Face posts of October 27 and 28 (NVIDIA's transcription model; AI2's "Can an LLM judge whether a research idea is novel?") and the Stanford HAI posts of October 19 and 21 (evaluating medical AI, tracking the labor market) were not read in full. Anthropic's October 27 post on drug development and the current model list and prices in its public API documentation involve a significant conflict of interest, so we will cover them separately after cross-checking.

What we could not confirm in these stories

Mistral Large 4: The Intelligence Index v5.0 score of 26 on the Artificial Analysis model page does not say whether it is for the open-weight version or the API version, so we could not settle that (Artificial Analysis alone). All company performance figures are company charts. "The quantized version runs on four GPUs" is from TNW alone, and we could not determine why "API pricing not published" (TNW) conflicts with the price listed in the official post. The model card has no release date, so we used the date on Mistral's news page (October 29).
Anthropic AI for Science: We could not confirm it with media beyond the company post. The gap between "20" and the 18 we counted, and the credit amount, duration and application method, are not in the post.
Epoch AI chip tally: The sources and assumptions behind the sales estimates were blank sections on the page, so we could not verify them. "Signs of deceleration that are not significant" relays Epoch's own wording.
Cursor notice: The user count ("at most 5,000") is a media citation, Anthropic's position was quoted by BankInfoSecurity alone, and the attacker's route in appears nowhere.

What we read alongside

We opened and checked sources of different kinds ourselves. They are grouped as company announcements, a model card, research organizations, an independent evaluator, a security firm, and press coverage.

Company announcements:

Model card:

huggingface.comistralai/Mistral-Large-4-1T-A52B (Hugging Face)English, parameters, architecture, context and license listings

Research organizations:

Independent evaluator:

Security firm:

sophos.comMalicious Claude Code extension distributed through Cursor marketplace (Sophos X-Ops)English, security firm summary

Press:

Our earlier articles: