AutoformBench
Agents fill in missing Lean proofs for mathematical statements and submit their solutions to a clean grading workspace.
The release includes 341 audited tasks and model results with task costs.
- Made by
- Meta FAIR
- Released
- October 2026
- Kind
- Formal proofs
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Written proof
- Tasks
- All public
- Measures
- Capability
- Subject
- Mathematics
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal, Code sandbox
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents
- Strongest model tested
- GPT-6 Astra #4 on the capability index
- Last model added
- September 2026
- Scores refreshed
- November 20th, 2026
GPT-6 Astra leads with 82.4% among models released by September 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- Claude Opus 5.5
- and 12 more
- 1 GPT-6 Astra source-present 82.4%
- 2 Claude Opus 5 source-present 61.9%
- 3 GPT-5.6 Sol source-present 58.9%
- 4 Claude Fable 5 source-present 53.1%
- 5
DeepSeek V4.1 Flash source-present 52.8%
- 6
Gemini 3.7 Flash source-present 52.5%
- 7 Kimi K3 source-present 48.4%
- 8 Claude Sonnet 5 source-present 44.6%
- 9 Claude Opus 4.8 source-present 44%
Source: Official AutoformBench leaderboard
| # | Model | Score | Setting | Date | Link | Source |
|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 82.4% | source-present | Sep 2026 | raw.githubusercontent.com | Official leaderboard |
| 2 | Claude Opus 5 | 61.9% | source-present | Jul 2026 | raw.githubusercontent.com | Official leaderboard |
| 3 | GPT-5.6 Sol | 58.9% | source-present | Jul 2026 | raw.githubusercontent.com | Official leaderboard |
| 4 | Claude Fable 5 | 53.1% | source-present | Jun 2026 | raw.githubusercontent.com | Official leaderboard |
| 5 | | 52.8% | source-present | Sep 2026 | raw.githubusercontent.com | Official leaderboard |
| 6 | | 52.5% | source-present | Aug 2026 | raw.githubusercontent.com | Official leaderboard |
| 7 | Kimi K3 | 48.4% | source-present | Jul 2026 | raw.githubusercontent.com | Official leaderboard |
| 8 | Claude Sonnet 5 | 44.6% | source-present | Jun 2026 | raw.githubusercontent.com | Official leaderboard |
| 9 | Claude Opus 4.8 | 44% | source-present | May 2026 | raw.githubusercontent.com | Official leaderboard |
Rankings by part
We read 24 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as algebraic_combinatorics, algebraic_geometry_i and algebraic_topology_i. GPT-6 Astra leads 16 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
algebraic_combinatorics
9 models we track have a score here.
- 1 Claude Opus 4.8 ↑8 100%
- 2 Claude Fable 5 ↑2 100%
- 3 GPT-5.6 Sol 100%
- 4 Kimi K3 ↑3 100%
- 5 Claude Opus 5 ↓3 100%
- 6 GPT-6 Astra ↓5 100%
- 7 Claude Sonnet 5 ↑1 50%
- 8
Gemini 3.7 Flash ↓2 50%
- 9
DeepSeek V4.1 Flash ↓4 50%
algebraic_geometry_i
9 models we track have a score here.
- 1 GPT-6 Astra 75%
- 2 Claude Fable 5 ↑2 12.5%
- 3 GPT-5.6 Sol 12.5%
- 4 Kimi K3 ↑3 12.5%
- 5 Claude Opus 5 ↓3 12.5%
- 6
Gemini 3.7 Flash 12.5%
- 7
DeepSeek V4.1 Flash ↓2 12.5%
- 8 Claude Opus 4.8 ↑1 0%
- 9 Claude Sonnet 5 ↓1 0%
algebraic_topology_i
9 models we track have a score here.
- 1 GPT-6 Astra 62.5%
- 2 Claude Fable 5 ↑2 31.3%
- 3 Claude Opus 5 ↓1 31.3%
- 4 GPT-5.6 Sol ↓1 25%
- 5 Claude Sonnet 5 ↑3 18.8%
- 6 Kimi K3 ↑1 18.8%
- 7
Gemini 3.7 Flash ↓1 18.8%
- 8
DeepSeek V4.1 Flash ↓3 18.8%
- 9 Claude Opus 4.8 12.5%
introduction_to_arithmetic_geometry
9 models we track have a score here.
- 1 GPT-6 Astra 76.9%
- 2 GPT-5.6 Sol ↑1 61.5%
- 3 Claude Opus 5 ↓1 61.5%
- 4
Gemini 3.7 Flash ↑2 53.8%
- 5 Claude Fable 5 ↓1 50%
- 6
DeepSeek V4.1 Flash ↓1 50%
- 7 Claude Opus 4.8 ↑2 46.2%
- 8 Kimi K3 ↓1 42.3%
- 9 Claude Sonnet 5 ↓1 38.5%
boolean_functions
9 models we track have a score here.
- 1 Claude Fable 5 ↑3 100%
- 2 GPT-5.6 Sol ↑1 100%
- 3 Claude Opus 5 ↓1 100%
- 4 GPT-6 Astra ↓3 100%
- 5
DeepSeek V4.1 Flash 100%
- 6 Claude Opus 4.8 ↑3 50%
- 7 Kimi K3 50%
- 8
Gemini 3.7 Flash ↓2 50%
- 9 Claude Sonnet 5 ↓1 0%
buildings
9 models we track have a score here.
- 1 GPT-6 Astra 100%
- 2 GPT-5.6 Sol ↑1 90.9%
- 3 Claude Opus 5 ↓1 81.8%
- 4
Gemini 3.7 Flash ↑2 81.8%
- 5 Claude Fable 5 ↓1 63.6%
- 6 Kimi K3 ↑1 63.6%
- 7
DeepSeek V4.1 Flash ↓2 63.6%
- 8 Claude Opus 4.8 ↑1 54.5%
- 9 Claude Sonnet 5 ↓1 45.5%
functions_of_a_complex_variable
9 models we track have a score here.
- 1 Claude Opus 5 ↑1 100%
- 2 GPT-6 Astra ↓1 100%
- 3 GPT-5.6 Sol 83.3%
- 4 Claude Fable 5 75%
- 5 Claude Sonnet 5 ↑3 75%
- 6 Kimi K3 ↑1 75%
- 7
DeepSeek V4.1 Flash ↓2 75%
- 8 Claude Opus 4.8 ↑1 66.7%
- 9
Gemini 3.7 Flash ↓3 66.7%
theory_of_computation
9 models we track have a score here.
- 1 GPT-6 Astra 87.5%
- 2 Claude Opus 4.8 ↑7 50%
- 3 Claude Fable 5 ↑1 50%
- 4 Claude Sonnet 5 ↑4 50%
- 5 GPT-5.6 Sol ↓2 50%
- 6 Kimi K3 ↑1 50%
- 7 Claude Opus 5 ↓5 50%
- 8
Gemini 3.7 Flash ↓2 50%
- 9
DeepSeek V4.1 Flash ↓4 50%
differential_analysis
9 models we track have a score here.
- 1 GPT-6 Astra 94.4%
- 2 Kimi K3 ↑5 77.8%
- 3 Claude Opus 5 ↓1 77.8%
- 4
DeepSeek V4.1 Flash ↑1 77.8%
- 5 GPT-5.6 Sol ↓2 72.2%
- 6 Claude Opus 4.8 ↑3 66.7%
- 7 Claude Fable 5 ↓3 66.7%
- 8 Claude Sonnet 5 66.7%
- 9
Gemini 3.7 Flash ↓3 55.6%
double_hecke_algebras
9 models we track have a score here.
- 1 GPT-6 Astra 80%
- 2 Claude Opus 5 40%
- 3 GPT-5.6 Sol 26.7%
- 4 Kimi K3 ↑3 26.7%
- 5 Claude Fable 5 ↓1 20%
- 6
DeepSeek V4.1 Flash ↓1 20%
- 7
Gemini 3.7 Flash ↓1 13.3%
- 8 Claude Opus 4.8 ↑1 6.67%
- 9 Claude Sonnet 5 ↓1 6.67%
elliptic_curves
9 models we track have a score here.
- 1 GPT-6 Astra 40.7%
- 2 GPT-5.6 Sol ↑1 25.9%
- 3 Claude Opus 5 ↓1 25.9%
- 4
Gemini 3.7 Flash ↑2 25.9%
- 5 Claude Fable 5 ↓1 18.5%
- 6 Kimi K3 ↑1 18.5%
- 7
DeepSeek V4.1 Flash ↓2 18.5%
- 8 Claude Opus 4.8 ↑1 14.8%
- 9 Claude Sonnet 5 ↓1 14.8%
graph_theory_and_additive_combinatorics
9 models we track have a score here.
- 1 GPT-6 Astra 66.7%
- 2 Claude Opus 4.8 ↑7 44.4%
- 3 Claude Fable 5 ↑1 44.4%
- 4 Claude Sonnet 5 ↑4 44.4%
- 5 GPT-5.6 Sol ↓2 44.4%
- 6 Claude Opus 5 ↓4 44.4%
- 7
Gemini 3.7 Flash ↓1 44.4%
- 8
DeepSeek V4.1 Flash ↓3 44.4%
- 9 Kimi K3 ↓2 22.2%
high_dimensional_statistics
9 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 100%
- 2 Claude Opus 5 100%
- 3 GPT-6 Astra ↓2 100%
- 4 Claude Fable 5 96%
- 5
DeepSeek V4.1 Flash 96%
- 6 Claude Sonnet 5 ↑2 88%
- 7 Kimi K3 88%
- 8
Gemini 3.7 Flash ↓2 88%
- 9 Claude Opus 4.8 84%
lie_groups_lie_algebras
9 models we track have a score here.
- 1 GPT-6 Astra 64.3%
- 2 GPT-5.6 Sol ↑1 42.9%
- 3 Claude Opus 5 ↓1 39.3%
- 4
DeepSeek V4.1 Flash ↑1 39.3%
- 5 Claude Fable 5 ↓1 35.7%
- 6 Claude Sonnet 5 ↑2 35.7%
- 7
Gemini 3.7 Flash ↓1 35.7%
- 8 Kimi K3 ↓1 32.1%
- 9 Claude Opus 4.8 28.6%
generalized_low_rank_models
9 models we track have a score here.
- 1 GPT-6 Astra 100%
- 2 Claude Fable 5 ↑2 80%
- 3 GPT-5.6 Sol 80%
- 4 Claude Opus 5 ↓2 80%
- 5
Gemini 3.7 Flash ↑1 80%
- 6 Claude Opus 4.8 ↑3 60%
- 7 Claude Sonnet 5 ↑1 60%
- 8 Kimi K3 ↓1 60%
- 9
DeepSeek V4.1 Flash ↓4 60%
matrix_concentration_inequalities
9 models we track have a score here.
- 1 GPT-6 Astra 100%
- 2 Claude Opus 4.8 ↑7 0%
- 3 Claude Fable 5 ↑1 0%
- 4 Claude Sonnet 5 ↑4 0%
- 5 GPT-5.6 Sol ↓2 0%
- 6 Kimi K3 ↑1 0%
- 7 Claude Opus 5 ↓5 0%
- 8
Gemini 3.7 Flash ↓2 0%
- 9
DeepSeek V4.1 Flash ↓4 0%
noncommutative_algebra
9 models we track have a score here.
- 1 GPT-6 Astra 77.3%
- 2 GPT-5.6 Sol ↑1 63.6%
- 3 Claude Opus 5 ↓1 63.6%
- 4
DeepSeek V4.1 Flash ↑1 59.1%
- 5
Gemini 3.7 Flash ↑1 54.5%
- 6 Claude Fable 5 ↓2 50%
- 7 Claude Sonnet 5 ↑1 45.5%
- 8 Kimi K3 ↓1 45.5%
- 9 Claude Opus 4.8 40.9%
number_theory_i
9 models we track have a score here.
- 1 GPT-6 Astra 75%
- 2 GPT-5.6 Sol ↑1 33.3%
- 3 Claude Opus 5 ↓1 33.3%
- 4 Claude Fable 5 20.8%
- 5 Kimi K3 ↑2 20.8%
- 6
Gemini 3.7 Flash 20.8%
- 7
DeepSeek V4.1 Flash ↓2 20.8%
- 8 Claude Opus 4.8 ↑1 12.5%
- 9 Claude Sonnet 5 ↓1 12.5%
introduction_to_pdes
9 models we track have a score here.
- 1 Claude Opus 5 ↑1 100%
- 2 GPT-6 Astra ↓1 100%
- 3 Claude Fable 5 ↑1 87.5%
- 4 GPT-5.6 Sol ↓1 87.5%
- 5 Kimi K3 ↑2 87.5%
- 6
Gemini 3.7 Flash 87.5%
- 7 Claude Opus 4.8 ↑2 75%
- 8 Claude Sonnet 5 75%
- 9
DeepSeek V4.1 Flash ↓4 75%
theory_of_probability
9 models we track have a score here.
- 1 Claude Fable 5 ↑3 100%
- 2 Kimi K3 ↑5 100%
- 3 Claude Opus 5 ↓1 100%
- 4
Gemini 3.7 Flash ↑2 100%
- 5 GPT-6 Astra ↓4 100%
- 6
DeepSeek V4.1 Flash ↓1 100%
- 7 Claude Sonnet 5 ↑1 75%
- 8 GPT-5.6 Sol ↓5 75%
- 9 Claude Opus 4.8 50%
optimizing_with_sparsity_inducing_penalties
9 models we track have a score here.
- 1 Claude Fable 5 ↑3 100%
- 2 GPT-5.6 Sol ↑1 100%
- 3 Claude Opus 5 ↓1 100%
- 4
Gemini 3.7 Flash ↑2 100%
- 5 GPT-6 Astra ↓4 100%
- 6
DeepSeek V4.1 Flash ↓1 100%
- 7 Claude Opus 4.8 ↑2 87.5%
- 8 Claude Sonnet 5 87.5%
- 9 Kimi K3 ↓2 87.5%
stacks_project
9 models we track have a score here.
- 1 GPT-6 Astra 96.9%
- 2 Claude Opus 5 87.5%
- 3
Gemini 3.7 Flash ↑3 84.4%
- 4 Claude Fable 5 75%
- 5 GPT-5.6 Sol ↓2 75%
- 6
DeepSeek V4.1 Flash ↓1 68.8%
- 7 Claude Opus 4.8 ↑2 62.5%
- 8 Claude Sonnet 5 62.5%
- 9 Kimi K3 ↓2 62.5%
learning_with_submodular_functions
9 models we track have a score here.
- 1 Claude Opus 5 ↑1 100%
- 2 GPT-6 Astra ↓1 100%
- 3 GPT-5.6 Sol 93.3%
- 4
DeepSeek V4.1 Flash ↑1 93.3%
- 5 Claude Opus 4.8 ↑4 86.7%
- 6 Claude Fable 5 ↓2 86.7%
- 7
Gemini 3.7 Flash ↓1 86.7%
- 8 Kimi K3 ↓1 80%
- 9 Claude Sonnet 5 ↓1 73.3%
tensor_categories
9 models we track have a score here.
- 1 GPT-6 Astra 100%
- 2 GPT-5.6 Sol ↑1 38.5%
- 3 Claude Fable 5 ↑1 30.8%
- 4 Claude Sonnet 5 ↑4 30.8%
- 5 Claude Opus 5 ↓3 30.8%
- 6
DeepSeek V4.1 Flash ↓1 30.8%
- 7 Kimi K3 23.1%
- 8
Gemini 3.7 Flash ↓2 23.1%
- 9 Claude Opus 4.8 15.4%
The best score, 82.4%, costs $43 a task
Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official AutoformBench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| GPT-6 Astra | source-present | 82.4% | $43 | 36,958 | |
| Claude Opus 5 | source-present | 61.9% | $44 | 639,377 | |
| GPT-5.6 Sol | source-present | 58.9% | $44 | 60,790 | |
| Claude Fable 5 | source-present | 53.1% | $97 | 420,444 | |
| source-present | 52.8% | $12 | 8,223,154 | ||
| source-present | 52.5% | $17 | 380,530 | ||
| Kimi K3 | source-present | 48.4% | $42 | 343,017 | |
| Claude Sonnet 5 | source-present | 44.6% | $42 | 1,094,963 | |
| Claude Opus 4.8 | source-present | 44% | $27 | 396,627 |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official AutoformBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official AutoformBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Math
- AIME 2026 In use Models solve the 30 integer-answer problems from the 2026 AIME I and II math competitions. Scored by exact answer match. GPT-5.5 100%
- IMO-AnswerBench In use 400 olympiad problems with short verifiable answers, robustified against memorization. Scored by answer match. GLM-5.2 91%
- HMMT 2026 In use Models solve the 33 final-answer problems from the February 2026 Harvard-MIT Mathematics Tournament. Scored by exact answer match. GPT-5.5 98.5%
- PolyMath In use Models solve 9,000 parallel math problems in 18 languages across four difficulty levels, from school math to frontier math. Final answers are checked.
Qwen3.7-Max 86.5%
- MathArena (Apex, IMO, USAMO) In use Evaluates models on newly released math competitions after release. Final-answer or human/LLM-judged proof scores. Claude Opus 4.8 81.3%
- MathArena Apex Shortlist In use Models solve 47 hard final-answer problems that MathArena selected from 2025 competitions, where the best models scored about 50% at selection. Scored by accuracy over repeated attempts. GPT-5.5 98.4%
Data audit
All 10 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 4 pass 1 note
- Pass Independent runs Scores from the official leaderboard (9 entries).
- Pass Models scored 9 models we track.
- Pass Newest score The newest is from September 2026, for DeepSeek V4.1 Flash.
- Note Frontier models Only 1 of the top 10 has a score. The strongest tested is GPT-6 Astra (#4).
- Pass Headroom The best is 82.4%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on github.com: 9 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.