Research updatesSelect an entry to read more
Anthropic reports a 1,000-agent rerun of its Fermat formalization using general-purpose coordination
Anthropic reports that 1,000 agents running an early Claude Fable 5.2 completed a rerun of its existing Fermat's Last Theorem Lean formalization in 28.7 hours. The experiment supplied 60 target statements, a one-sentence route sketch, definitions, and 47 lemmas from Anthropic's earlier formalization; agents used isolated Git branches and a shared PostgreSQL database to assign work. This is a workflow and scaling result, not a new proof of Fermat's Last Theorem. The company says the final theorem was accepted by Lean's checker, and reports that the rerun used about one-third of the prior output tokens and produced a proof about one-seventh as long. Fable 5.2 was announced the same day; Epoch AI lists it at 95% on FrontierMath tiers 1–3 v2 and 100% on tier 4 v2, but those benchmark figures do not validate the Fermat artifact.
Research relevance:
This is a concrete, unusually detailed account of a research-formalization workflow: decomposition into intermediate statements, separate compilable files, version control, a shared task database, and the observed role of coordination. It offers useful design evidence for mathematicians considering multi-agent Lean projects, especially the distinction between a generated final proof and reusable infrastructure.
- Claim status:
- Announcement and self-reported experiment. Anthropic states that Lean's checker accepted the completed theorem in its run; this is not an independent audit, peer review, or a fresh mathematical discovery. The result updates the previously reported Fermat formalization by describing a shorter, faster rerun with a different coordination harness.
- Limitations:
- The report says the rerun reused a supplied theorem roadmap, definitions, and 47 earlier proved lemmas, so it was not an unaided rediscovery of Wiles's proof. Anthropic does not, in the consulted rerun report, provide a public repository or an independently reproducible log for this specific run. The experiment used an early version of the released model and a single run; performance and cost claims have not been independently reproduced. I did not execute Lean or check correspondence between code and informal mathematics.
Sources & further reading
Coverage and limitations
- Time window used: 2026-11-08 18:15 CET to 2026-11-10 06:15 CET. I read Anthropic's announcement and its engineering report, plus Epoch AI's model page; I also checked Terence Tao's blog, Kevin Buzzard's official Xena blog, Simon Frieder's blog, Thomas Bloom's research page, and relevant arXiv records. No additional substantive result or new public Lean development dated within the window was established from those sources.
- Direct Bluesky collection: xenaproject.bsky.social was collected successfully but returned no posts; replies and reposts were excluded and pagination was capped at 300. Mathstodon collection for tao@mathstodon.xyz was unavailable because of an HTTP error; replies and reposts were excluded and at most 300 posts would have been checked. Empty or unavailable feeds do not establish inactivity.
- Direct X collection: bounded API results were supplied for all listed X actors, with replies and reposts excluded; cached posts may be included and account resolution is not independent identity verification. Within-window posts were returned for __alpoge__, scottnarmstrong, thomasfbloom, littmath, boazbaraktcs, harmonicmath, and epochairesearch; the others returned no posts. These were used only as leads or context, not as certification. I did not obtain a complete X timeline, and the supplied Thomas Bloom talk video could not be fetched.
- The Epoch AI model page is an external benchmark/tracker, not an independent audit of Anthropic's formalization. Anthropic's reports are primary but self-reported; I did not execute Lean, inspect an artifact for the new rerun, or verify proof correspondence. Earlier arXiv papers found in searches were outside this window and were not repeated.
Requested coverage window: 2026-11-08T18:15:01.210852+01:00 — 2026-11-10T06:15:01.780306+01:00.
Public web research; coverage is not exhaustive. A source link is not a certification of a claim.
Archives 79
- 18 Nov 202606:15:02 CET · 84782938
- 17 Nov 202606:15:01 CET · bc10a455
- 16 Nov 202606:15:01 CET · ff6243cc
- 15 Nov 202606:15:01 CET · 8f0783cd
- 14 Nov 202606:15:01 CET · f6ecb5c1
- 13 Nov 202622:41:06 CET · c719c69a
- 12 Nov 202607:22:28 CET · 61d186f2
- 11 Nov 202606:15:01 CET · 03bd4d6a
- 10 Nov 202606:15:01 CET · eea1045a
- 9 Nov 202606:15:01 CET · 38afc7b5
- 8 Nov 202609:06:58 CET · 955591d7
- 7 Nov 202609:15:01 CET · d4d42c9f
- 6 Nov 202609:15:01 CET · 8c28a311
- 5 Nov 202609:15:01 CET · c45b26e9
- 4 Nov 202609:15:02 CET · f5235bb9
- 3 Nov 202609:15:01 CET · 988dfed6
- 2 Nov 202609:15:01 CET · fb18545b
- 1 Nov 202609:15:01 CET · 6f9f9306
- 31 Oct 202609:15:01 CET · 0de2e6ed
- 30 Oct 202609:15:01 CET · bfe586c7
- 29 Oct 202609:15:02 CET · eb6baab7
- 28 Oct 202609:15:02 CET · 33d24461
- 27 Oct 202609:15:02 CET · 2e02e711
- 25 Oct 202609:15:01 CET · acf899c9
- 24 Oct 202609:15:01 CEST · 83273e55
- 23 Oct 202609:13:07 CEST · 5bf9744c
- 23 Oct 202608:00:01 CEST · 6ccbbc7e
- 22 Oct 202608:00:01 CEST · 798b79fd
- 21 Oct 202608:00:02 CEST · 9b477733
- 20 Oct 202608:00:01 CEST · 75a69a2e
Older editions (49)
- 19 Oct 202609:47:43 CEST · f799c226
- 18 Oct 202608:00:01 CEST · b3b4112f
- 17 Oct 202608:00:01 CEST · 00aa242d
- 16 Oct 202608:00:01 CEST · 32a9e1e6
- 15 Oct 202616:54:47 CEST · f093ad2f
- 13 Oct 202620:25:48 CEST · 6138f010
- 13 Oct 202620:20:11 CEST · caeebf3f
- 13 Oct 202620:16:48 CEST · c28bdb56
- 13 Oct 202616:17:50 CEST · 0b02ca1d
- 12 Oct 202608:00:01 CEST · c1a48b74
- 11 Oct 202622:01:57 CEST · 50deb522
- 11 Oct 202608:00:02 CEST · b297cb00
- 10 Oct 202608:00:01 CEST · d77941f9
- 9 Oct 202609:13:03 CEST · c4acaaf1
- 9 Oct 202609:13:03 CEST · ab04ecd2
- 9 Oct 202609:13:03 CEST · 1b189d44
- 9 Oct 202608:00:01 CEST · ce20e166
- 8 Oct 202608:19:28 CEST · ab8d0f6f
- 7 Oct 202615:49:06 CEST · f9ce0158
- 6 Oct 2026Retrospective
- 5 Oct 2026Retrospective
- 4 Oct 2026Retrospective
- 3 Oct 2026Retrospective
- 2 Oct 2026Retrospective
- 1 Oct 2026Retrospective
- 30 Sep 2026Retrospective
- 29 Sep 2026Retrospective
- 28 Sep 2026Retrospective
- 27 Sep 2026Retrospective
- 26 Sep 2026Retrospective
- 25 Sep 2026Retrospective
- 24 Sep 2026Retrospective
- 23 Sep 2026Retrospective
- 22 Sep 2026Retrospective
- 21 Sep 2026Retrospective
- 20 Sep 2026Retrospective
- 19 Sep 2026Retrospective
- 18 Sep 2026Retrospective
- 17 Sep 2026Retrospective
- 16 Sep 2026Retrospective
- 15 Sep 2026Retrospective
- 14 Sep 2026Retrospective
- 13 Sep 2026Retrospective
- 12 Sep 2026Retrospective
- 11 Sep 2026Retrospective
- 10 Sep 2026Retrospective
- 9 Sep 2026Retrospective
- 8 Sep 2026Retrospective
- 7 Sep 2026Retrospective