OpenAI released 722 math manuscripts in 372 result families. Its readme says the model received about 4,000 problems, but the release does not say how many produced no manuscript. We can see the scale of the output, but we cannot calculate a success rate from it.
That distinction matters throughout the release. A manuscript, a formally checked result, and an answer to the problem originally posed are different things.
722 Manuscripts, but No Success Rate
OpenAI reports an average of about 3 hours of ChatGPT Pro thinking compute per result. One report says the main results of 162 manuscripts have been formalized in Lean, software that checks the steps of a formally written proof.
Those numbers describe a large amount of work. They do not tell us how many of the roughly 4,000 problems led to no manuscript, or how many of the remaining manuscripts would hold up to the same kind of check. OpenAI also says some unformalized results could have issues.
A Lean Check Has a Boundary
Lean checks the formal statement someone entered and the proof steps that follow it. It does not check whether that statement matches the original problem, or whether a person understands the proof. So the exact statement deserves attention even when a check passes.
The quasi Riemann hypothesis entry shows why. It concerns Hecke L-functions over one number field and reports no zeros where the real part of S is above 7/8. The full Riemann hypothesis would put the line at 1/2, and this entry does not cover the Riemann zeta function itself. Its name is easy to read more broadly than its stated result.
Check status also differs from entry to entry. OpenAI marks its Catalan’s constant result, the two-point Chala result, and a result disproving Zariski cancellation in dimension four as Lean checked. Its list shows no Lean check for entries on Hilbert’s 10th problem over the rational numbers or a formula in the Birch and Swinnerton-Dyer problem.
The useful reading habit is straightforward: look at the claim itself, then look at its check status. A Lean check tells you something important about a formal proof. It does not, by itself, settle what the original problem asked.
Reading the Work Takes People
Checking a proof and reproducing the process that produced it are separate jobs. OpenAI says almost every result came from one prompt to one AI agent. It disclosed the average compute time and 10 reasoning summaries, but has not named the model or released the prompts. An advisory group recommended disclosing both.
There is a human workload, too. Quanta reported researchers posting preprints early to avoid being scooped by AI, declining Math Overflow activity, and limits on arXiv submissions. Terence Tao called for a slower pace and has emphasized that even proofs passing Lean still need a person to read them.
Hundreds of manuscripts can arrive together. Reading what each one actually claims, and understanding a proof even after its formal steps pass, takes time. That is why the scale of this release makes the difference between a checked statement and human review especially important.
Coding Agents Are Taking On Longer Tasks
The script’s other thread concerns agents that keep working across coding tasks. Anthropic’s Claude Code Projects lets one ongoing conversation coordinate separate task threads. Each thread has its own cloud session and branch, and can run on the user’s computer when it needs local files or tools.
A thread can open and watch a pull request, push fixes after CI failures, and address review comments. In Anthropic’s example walkthrough, a planned daily check of real user metrics opens an investigation thread if it finds a regression. Anthropic says Projects opened to every Pro and Max user on the waitlist; the walkthrough calls it a gradual beta rollout. Several full Claude Code sessions running together also use plan limits faster than one.
OpenAI made Codex steering instant and released GPT-6.1 Sol Ultrafast across the API, Codex, and ChatGPT Work. OpenAI describes the model as up to 8x faster than Sol Standard. Codex CLI 0.162.0 added managed Git worktrees and pinnable tasks, while Codex Cloud can now reach resources on a Tailscale tailnet.
Anthropic also cut the Sonnet 5.5 API cache-read price to $0.10 per million tokens. Anthropic describes that as about 20% cheaper on most agent runs; Claude Code usage limits did not change. These are specific speed and cost descriptions, not measurements of every workflow.
For the math release, independent reviewers still lack the model name and prompts needed to reproduce OpenAI’s reported single-agent process.