BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

ProofBench

Full written proofs graded for rigour (Vals AI).

As of 24 Oct 2026, Claude Opus 5.5 leads ProofBench on BenchLeader with 100.0%, ahead of Claude Fable 5.1 at 100.0%, across 78 model configurations with a published result.

Published by
Vals AIdata via Epoch AI Benchmarking Hub
Category
Maths
Index weight
0.5
Models
78
Data as of
24 Oct 2026
Newest result
12 Oct 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

ProofBench asks for complete written proofs of competition and undergraduate theorems, not just final answers, and grades them for rigour and completeness.

How it is scored

Percent of full credit across proofs, graded by a rubric with human review, published by Vals AI.

What to keep in mind

Proof grading involves judgment, so the grading protocol matters as much as the model. Scores are far below final-answer maths benchmarks and should not be compared to them.

Top 15
  1. 1Claude Opus 5.5 (max)100.0%
  2. 2Claude Fable 5.1 (max)100.0%
  3. 3Claude Sonnet 5.5 (max)100.0%
  4. 4Claude Opus 5 (max)99.0%
  5. 5GPT-6 Astra99.0%
  6. 6Gemini 4 Argon99.0%
  7. 7GPT-6.1 Sol99.0%
  8. 8Claude Fable 5 (max)95.0%
  9. 9Claude Haiku 5.5 (max)89.0%
  10. 10Kimi K387.0%
  11. 11Muse Wonder84.0%
  12. 12GPT-5.6 Sol (max)83.0%
  13. 13GPT-6 Sol83.0%
  14. 14Claude Sonnet 5 (max)77.0%
  15. 15GPT-5.6 Terra (xhigh)74.0%
78 of 78
#
1Claude Opus 5.5max100.0%72.62026-09-22
2Claude Fable 5.1max100.0%70.52026-09-01
3Claude Sonnet 5.5max100.0%69.32026-09-28
4Claude Opus 5max99.0%68.82026-07-24
5GPT-6 Astra99.0%68.02026-09-03
6Gemini 4 Argon99.0%64.92026-09-30
7GPT-6.1 Sol99.0%63.02026-09-29
8Claude Fable 5max95.0%69.32026-06-09
9Claude Haiku 5.5max89.0%60.52026-10-07
10Kimi K387.0%59.12026-07-16
11Muse Wonder84.0%63.72026-10-12
12GPT-5.6 Solmax83.0%68.02026-07-09
13GPT-6 Sol83.0%–2026-09-22
14Claude Sonnet 5max77.0%58.42026-06-30
15GPT-5.6 Terraxhigh74.0%62.82026-07-09
16MiMo-V2.6-Pro70.0%64.02026-09-21
17Claude Opus 4.8max69.0%63.32026-05-28
18GPT-6 Luna64.0%–2026-09-22
19MiMo-V2.6-Flash63.0%59.72026-09-21
20GPT-5.6 Lunamax60.0%59.92026-07-09
21Muse Spark 1.3max58.0%67.12026-09-02
22Qwen3.8 Maxmax58.0%63.32026-08-02
23Gemini 3.7 Flash58.0%53.42026-08-13
24GPT-5.4xhigh56.0%64.72026-03-05
25DeepSeek V4 Flash56.0%52.22026-07-31
26Muse Spark 1.355.0%59.82026-09-02
27Claude Opus 4.7max54.0%62.42026-04-16
28DeepSeek V4.1 Flash54.0%–2026-09-09
29Grok 4.651.0%57.82026-08-12
30GPT-5.5xhigh50.0%66.12026-04-23
31Claude Opus 4.6max50.0%57.42026-02-05
32DeepSeek V4 Pro50.0%53.52026-08-13
33GLM 5.3max49.0%63.52026-08-14
34Gemini 3.8 Flash48.0%55.72026-09-02
35Claude Sonnet 4.6max45.0%56.02026-02-17
36Muse Spark 1.243.0%–2026-08-05
37Step 5 Preview42.0%60.42026-09-18
38Muse Spark 1.139.0%62.52026-07-09
39Claude Opus 4.536.0%57.92025-11-24
40Gemini 3.6 Flash36.0%49.52026-07-21
41GLM 5.2max35.0%62.32026-06-16
42Gemini 3.5 Flashhigh31.0%63.12026-05-19
43Grok 4.5high31.0%59.32026-07-08
44Gemini 3.1 Pro26.0%60.62026-02-19
45Qwen3.7max26.0%60.62026-05-19
46Grok 4.726.0%–2026-09-21
47GLM 5.122.2%55.92026-04-07
48MiMo-V2.5-Pro22.0%54.9–
49GPT-5.4 Minixhigh21.0%56.32026-03-17
50GLM 5.3 Flashmax21.0%54.22026-08-20
51Gemini 3 Pro20.0%60.42025-11-18
52Claude Sonnet 4.519.0%53.72025-09-29
53GPT-5high18.0%57.92025-08-07
54MiniMax M318.0%56.42026-06-01
55Muse Spark17.0%64.32026-04-08
56DeepSeek V4 Promax16.0%60.82026-04-24
57Kimi K2.616.0%55.42026-04-20
58Qwen3.8 27B16.0%54.62026-08-14
59GPT-5.2xhigh15.0%60.72025-12-11
60Gemini 3 Flash15.0%57.02025-12-17

More maths benchmarks · All LLM benchmarks

Cite as: BenchLeader, “ProofBench leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/proofbench.

ProofBench: questions

What does ProofBench measure?
ProofBench asks for complete written proofs of competition and undergraduate theorems, not just final answers, and grades them for rigour and completeness. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads ProofBench?
Claude Opus 5.5 leads ProofBench with 100.0% as of 24 Oct 2026, ahead of Claude Fable 5.1 at 100.0%.
How many models have ProofBench results?
78 model configurations have a ProofBench result on BenchLeader, all taken from Vals AI via Epoch AI Benchmarking Hub.
Who runs ProofBench and how often is it updated?
ProofBench is published by Vals AI. BenchLeader re-reads the published results once a day and records the date each result was published.
Does ProofBench count toward the BenchLeader Index?
Yes. ProofBench contributes to the maths category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.