The Signal — October 7, 2026

OpenAI put hundreds of math papers written by an unreleased model on GitHub and told readers up front that some of them may be wrong. Mistral shipped its biggest model as a guarded preview while it decides how to release the weights, and Anthropic laid out which defenders can use Claude with fewer cyber safeguards and how much those safeguards still block.

OpenAI posted 722 math manuscripts from an unreleased model, and says some unformalized results could have issues

On October 6 OpenAI published a public GitHub repository of mathematical papers and proof files produced by an internal model it has not released. OpenAI's announcement links that model to the system behind its Navier–Stokes claim in September. The repository's catalogue lists 722 manuscripts grouped into 372 "families," where a family bundles a main result with companion arguments, consequences or alternative proofs. Many of the proofs come with formal versions in Lean, a language that lets a computer check each step, though OpenAI says not all of them do yet. The release also includes ten abridged summaries of the model's reasoning on results such as the irrationality exponent of π and Kaplansky's direct-finiteness conjecture in characteristic two.

OpenAI also described how the model produced the results. It was given roughly 4,000 problems, and the average result used about three hours of ChatGPT Pro thinking compute. Most results came from one fixed procedure; the exceptions OpenAI names are a zero-free region for the Riemann zeta function, whose writeup a human edited for readability, and a proof of the Hodge Conjecture for CM abelian varieties. The README is blunt about status. It says the collection "includes results at different stages of verification" and that "some of the unformalized results could have issues," with corrections to be logged as new versions.

The release format follows advice from the Advisory Group on Mathematics and Artificial Intelligence, an unpaid panel of mathematicians at the Institute for Advanced Study that OpenAI consulted. In late September the group asked AI labs to publish results promptly, through academic channels where possible, to disclose the model, prompts and compute, and to stop using math results as marketing for their models. According to The Verge, the group says the batch includes solutions to "hundreds" of open questions. The group advised on how to release the work, and no outside reviewer has yet checked all 372 families. Sorting out which results are new, which are correct and which were already known will take mathematicians time, as The Verge notes. OpenAI says it is working toward releasing the model and will fund workshops on AI-produced results.

Sources: OpenAI · openai/math on GitHub · The Verge · AGMAI recommendations

Mistral Large 4 arrived as an API preview, with open weights promised for the end of the month

Mistral released Mistral Large 4, nicknamed Le Chonk, as a public preview through its own API on October 6. It is a multimodal mixture-of-experts model with roughly a trillion parameters, of which only a fraction run for any given token. Mistral's own pages disagree on the exact active count, so we are leaving that number out. The weights are not out yet. Mistral's Hugging Face page is a placeholder dated October 31 that promises open weights at the end of the month, and TechCrunch reports the model is reachable only through a guarded endpoint until safety testing is done. Mistral VP of Science Pierre Stock told TechCrunch the company will work with "trusted partners and governments" in the meantime so the weights can be used for defense rather than attacks.

Stock also said the model was trained entirely on Mistral's own compute, about 4,000 NVIDIA GPUs, and named cybersecurity, finance and chip design as its target uses. Chip design is core business for two of Mistral's main backers, ASML and Samsung. The first outside numbers are decent. Artificial Analysis scores the preview 38 on its Intelligence Index against a median of 26 for comparably priced models and lists prices of $1.36 per million input tokens and $4.18 per million output tokens. It also flags the model as very verbose: the preview generated about 200 million tokens on the index, where the median model used 81 million. Simon Willison, who tried the API, notes that the score of 38 sits just behind DeepSeek 4.1 Flash and is a large jump from Mistral Large 3's score of 9. He puts the model "maybe about 6 months behind the frontier."

The license for the coming weights has not been published, and Artificial Analysis currently lists the preview as proprietary. Whether the model ships with open weights, and on what terms, depends on what Mistral publishes on October 31.

Sources: Mistral on Hugging Face · TechCrunch · Artificial Analysis · Simon Willison

Anthropic split its cyber access program into three tiers, and its own test shows how often safeguards stop the model

Anthropic says its generally available Claude models block most cyber work by default. Since spring the company has let some defenders around those blocks through two separate programs: Project Glasswing, which gave a group of critical-software organizations access to its Mythos models, and the Cyber Verification Program, which loosened safeguards on Opus and Sonnet for vetted security teams. On October 6 Anthropic merged the two into one program with three tiers, each covering Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models.

Defense Access covers incident response, malware reverse engineering and vulnerability analysis, and is open to security teams at companies, universities and government bodies, critical-infrastructure operators of any size, small security firms, open-source maintainers and individual researchers with a record of reported vulnerabilities. Anthropic says it aims to answer those applications within a few days. Red Team Access adds authorized penetration testing for organizations only and still blocks actions such as deploying ransomware or damaging physical systems. Applications take a few weeks to review. Specialized Access, with the fewest blocks, is reserved for organizations authorized to test systems like power grids, flight software and interbank transfers, and Anthropic vets each organization together with the US government. Existing Glasswing members move into that tier. Most enrolled organizations must let Anthropic retain their data so it can watch for misuse, at least until the zero-retention option the company has promised for later this fall arrives.

Anthropic ran Claude Opus 5.5 through its CyScenarioBench evaluation of multi-stage cyber operations, five attempts at each of ten challenges per tier. Without program access every task was blocked on the first prompt. In the Defense tier, 46 of 50 trials hit a block at some point. In the Red Team tier nothing was blocked and the model completed 34 of 50, which Anthropic says matches its success rate with no safeguards at all. Those are Anthropic's own tests. It also says Glasswing partners found at least 129,000 verified vulnerabilities between April and July, a figure Reuters repeats with Anthropic's caveat that it comes from a survey of only some partners.

Sources: Anthropic · Reuters · Quartz

On the Editor's Desk

South Korea's plan to put 4.7 trillion won into a homegrown frontier model got fresh coverage, but the amount was first reported in September and the money still needs approval from the National Assembly, so we are waiting for the budget vote. We also held the 18-month sentence for Michael Smith, who pleaded guilty to using bots to stream AI-generated songs for royalties, because we could not read the Justice Department's statement and only had secondhand reports. Google's EmbeddingGemma 2 looks useful for developers, but every benchmark so far is Google's own.