# A machine that proves Fermat in 29 hours, and the behemoth that's back (slim) > Vlad Cealicu on Anthropic's post about 1,000 agents on an early Claude Fable 5.2 reproducing the Lean proof of Fermat's Last Theorem in 28.7 hours. What the general tools got right, six decisions he'd make differently (the steering, the weaker monitor, zero blocks, one lead, an unassessed model at scale, one shared database role), how Anchor Terminal runs its own 50 agents and one coordinator, and why Google, the company whose researchers wrote the transformer paper, is back in the frontier fight and winning distribution. - Full: https://www.anchorterminal.com/blog/a-machine-that-proves-fermat.md (~3,400 tokens) · this version ~1,280 tokens · JSON https://www.anchorterminal.com/blog/a-machine-that-proves-fermat.json · canonical https://www.anchorterminal.com/blog/a-machine-that-proves-fermat - Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-13 - Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data) By Vlad Cealicu*, 13 November 2026. 1,610 words, about 7 min to read. - 1,000: agents on an early Claude Fable 5.2, one lead and 999 helpers - 28.7 hours: to the final theorem, against 11 days for the first formalisation in September - 1 million+: tool calls, and none of them blocked - 19: automatic reminders to the lead to keep going The run, from Anthropic's post. Anthropic's post on 9 November is about a team of 1,000 agents running an early version of Claude Fable 5.2 that reproduced the Lean proof of Fermat's Last Theorem in 28.7 hours, with a third of the output tokens and a proof a seventh as long as the one they formalised in September, which took 11 days, dozens of agents, a researcher and a platform built for the job called Prove2Me [[1]](#ref-1). ## What they got right The tools were a shell, a file editor, Lean with Mathlib, a Git branch for each agent with merge requests, and one shared Postgres database for coordination, and the whole point was to see whether general tools are enough [[1]](#ref-1). I've said for a while that I don't think MCP is the future and that plain, easy APIs win, and I use Postgres everywhere I need a database, so a post where Postgres and Git beat a domain platform is a post I was always going to like (I know, very biased). ## How we do it We run something with the same shape. The thing our runs taught us that matters here is that a rule in a prompt isn't a control. ## Six things I'd do differently The first is the comparison. Hours from the start of the run | Who | Time | Outcome | | --- | --- | --- | | The lead sets aside 47 of 60 statements | about 27 minutes | 3 proved, nearly every helper moved to the other 10 | | Activity stops, the run restarts from its last save | about hour 18 | 12 more hours given | | Langlands-Tunnell, the hardest statement, is proved | hour 28.4 | its captains had put the odds under 1% | | The final theorem is proved | hour 28.7 | 54 of 60 by hour 30 | | The September formalisation | 11 days | with Prove2Me, dozens of agents and a researcher | Time on a log scale. The second is the monitor. The third is the zero. The fourth is the lead. The fifth is the model. The sixth is the database. ## Meanwhile, Google While Anthropic was writing up a thousand agents on Fermat, Google was shipping. The transformer paper that all of this runs on was written in 2017 by researchers at Google Brain and Google Research [[9]](#ref-9). ## What I'd want from the next one I'd want the same run with the chain-of-thought monitor on from the start, a role per agent in the database, a lead that can be replaced without stopping the team, and the comparison made against a run that got the same 59 milestones and the same reminders. --- More in [Opinion](https://www.anchorterminal.com/blog/opinion/index.md) - Older: [I don't think you need Claude's harness](https://www.anchorterminal.com/blog/another-harness-for-business.md) - All posts: https://www.anchorterminal.com/blog/index.md Anchor Terminal grades the tools AI agents run on, with a source for every score. ## References Read on 13 November 2026. Quotes are as published. [1] Anthropic, claude.dev blog, Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl and Lance Martin, 9 November 2026. Safely scaling multi-agent teams with Claude Fable 5.2. Source for the run (1,000 agents, 28.7 hours, a third of the output tokens, a proof a seventh as long, the September formalisation's 11 days), the tools, VMs, tables and merge job, the sandbox and classifier, the starting materials, the 19 reminders, the restart near hour 18, the remark on human encouragement, the red-team results and the monitor used, the million tool calls and the replay, what a block on the lead does, the 27-minute triage, the timings for Langlands-Tunnell and the final theorem and the captains' odds, the quote on the alignment assessment, and the database role, limits and notes board. [2] Anchor Terminal, how our listing research runs, November 2026, not published. [3] Anchor Terminal research rules, November 2026, not published. [4] Anchor Terminal, how reviews work. [5] Google, Gemini API changelog, read 13 November 2026. Source for Gemini 4 Argon going generally available on 13 October 2026. https://ai.google.dev/gemini-api/docs/changelog [6] Anchor Terminal, the Gemini 4 Carbon listing, with Google's announcement of 28 October and developer post of 10 November as its sources. [7] Google, Koray Kavukcuoglu, 12 November 2026. Introducing Gemini 4 Hydrogen. [8] Google DeepMind, Gemini 4 Hydrogen model card, November 2026. Source for Google's table against Claude Sonnet 5.5 and GPT-6.1 Terra, with Hydrogen ahead on 7 of 10. https://deepmind.google/models/model-cards/gemini-4-hydrogen/ [9] Vaswani, A.