OpenAI drops 372 math results, nearly all from one prompt

Share
OpenAI drops 372 math results, nearly all from one prompt

A dense news day across three fronts: OpenAI published a mountain of mathematics from an unreleased model, SpaceX reportedly wants the debt market to fund its Nvidia orders, and Anthropic restructured its cyber program around three tiers of trust.

OpenAI published 372 mathematical result families from an internal frontier model — and says nearly all of them came from a single prompt handed to a single AI agent. The release comprises 722 manuscripts touching hundreds of open questions, plus Lean formalizations that let a computer check the proofs. In a nod to the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, OpenAI also shipped process details it usually keeps quiet: ten summaries of the model's reasoning, compute estimates (the average result consumed roughly three hours of ChatGPT Pro thinking), and statistics on attempted problems. The headline artifact sits in the repository as preprint number 109: integer multiplication below n log n, which the authors say disproves the Schönhage–Strassen optimality conjecture — a benchmark of fast multiplication for decades. Skepticism is running ahead of celebration: MIT mathematician Andrew Sutherland told Scientific American to treat claims of one-shotting problems with a single agent as unverified and to "ask for receipts," and OpenAI published only average compute, not the prompts. Our take: every number here is self-reported and unreplicated, but if even a fraction survives review, this is the first large batch of research results where a lab's unreleased model — not its shipping product — is the protagonist.


SpaceX is reportedly seeking to raise $40 billion — about $10 billion in bank loans and $30 billion in investment-grade debt — to buy Nvidia chips, with Apollo leading the deal. The Financial Times broke the story, and Reuters and Bloomberg both carried it within hours, each citing FT; SpaceX, Apollo, Nvidia and Pimco have not confirmed it. If the number holds, it would be the largest AI-chip financing yet, and the structure is the story: the compute buildout is being funded by debt rather than equity, with the deal expected to close in 2027.


Anthropic folded Project Glasswing into an expanded, three-tier Cyber Verification Program that gives vetted security teams its most capable models with progressively fewer safety blocks. Defense Access covers incident response and malware analysis with a review turnaround of days; Red Team Access adds authorized offensive testing; the top tier — Specialized Access, for testing systems like flight operations and power grids — is vetted in collaboration with the US government, and existing Glasswing members move over without reapproval. All tiers get Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1. Anthropic says Glasswing partners found 129,000 verified vulnerabilities between April and July, and its own benchmark draws the tier line sharply: every cyber task was blocked on the first prompt for users without program access, while top-tier access completed 34 of 50 runs with zero blocks.

What to watch: whether any mathematician starts replicating the OpenAI results — and whether the SpaceX figure survives confirmation beyond FT's sources.

Three hours of ChatGPT Pro thinking per theorem, and the prompts stay private — would you trust a proof you can't audit? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Princeton's Mengdi Wang: LLMs haven't made a real scientific discovery yet

Princeton's Mengdi Wang: LLMs haven't made a real scientific discovery yet

The models that cracked competition math still can't find a new particle — and Princeton's Mengdi Wang thinks she knows why. Large models learn "the most likely" answer, and science lives in the long tail. That was the core of a keynote from Mengdi Wang, director of Princeton's Center for AI Innovation, at the Inclusion·Bund Conference in Shanghai on September 10: the training method behind today's LLMs has conquered math and coding but, in her words, still hasn't produced a genuinely new…

Ten Claude agents formalize a 122-year-old physics problem

Ten Claude agents formalize a 122-year-old physics problem

A proof no human wrote, a 320-billion-parameter release from an unknown Chinese lab, and OpenAI's billing system quietly starting to carry other labs' models — the overnight lane had a research flavor. Ten Claude Sonnet 5.5 agents spent roughly 15 hours and 1,270 messages producing a 17,895-line Lean proof that settles the N=7 case of the Thomson problem — J.J. Thomson's 1904 question of how seven charged points arrange themselves on a sphere to minimize their mutual repulsion.

Anthropic ships Sonnet 5.5 as cyber safeguards move down a tier

Anthropic ships Sonnet 5.5 as cyber safeguards move down a tier

Anthropic's second Claude 5.5 model lands with frontier-class benchmark deltas at a mid-tier price — and the safety machinery that used to sit only on the top model. Meanwhile a chipmaker buys into the financing layer of the AI buildout, and Congress gets its most detailed AI-safety bill yet, with no path to a vote. Anthropic released Claude Sonnet 5.5 on September 28, and the interesting part is not the score — it is that Sonnet 5.5 beats Anthropic's own top model on the hardest agentic…

Shanghai AI Lab's Intern-S2 helped close a 30-year conjecture

Shanghai AI Lab's Intern-S2 helped close a 30-year conjecture

An AI-assisted proof landed in a form mathematicians can actually audit — and a second team with different tools checked the work. The arXiv paper proving the codimension-one case of a conjecture Étienne Ghys posed in 1995 now carries an explicit declaration of AI use: its author, Jiesong Zhang, says he used an internal version of Shanghai AI Lab's Intern-S2 "to assist in developing the mathematical ideas and arguments" across multiple rounds of dialogue, and thanks Shanghai AI Lab and Bowen…

H company opens Holo4: a 27B agent that drives your desktop

H company opens Holo4: a 27B agent that drives your desktop

Two releases in the same morning frame the agent debate neatly: a small open-weight model that can take over a desktop cheaply, and a safety stack being sold on silicon because the model-level guardrails keep failing. H company released Holo4, a computer-use agent family in two open-weight sizes — 27B dense and 35B-A3B mixture-of-experts — alongside Holo4's weights in BF16, FP8, NVFP4 and 4-bit GGUF formats on Hugging Face.