Benchmark Almanac

MysteryMechanism

GPT-6 Astra 53.1% leads since September 2026
Not in model cards

Agents recover a hidden mathematical law from two noisy observations and a few experiments they choose. The submitted expression is run on private probes and passes if its error is under a noise-aware threshold.

Inputs are anonymous, with no domain context. A blinded audit records whether each solved trace named the source domain.

Made by
Vals AI
Released
September 2026
Kind
Abstract puzzles and logic
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Strongest model tested
Claude Fable 5.2 #1 on the capability index
Last model added
November 2026
Scores refreshed
November 12th, 2026
31
Models scored on its own leaderboard

GPT-6 Astra leads with 53.1%

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

31 models. Scroll sideways for the rest.

  1. 53.1 GPT-6 Astra
  2. 51.8 Claude Fable 5.2
  3. 49.5 Claude Opus 5.5
  4. 49.1 Claude Sonnet 5.5
  5. 47.8 Claude Fable 5.1
  6. 46.4 GPT-6.1 Sol
  7. 45.5 Gemini 4 Argon
  8. 37.4 Claude Opus 5
  9. 36.5 Gemini 3.8 Flash
  10. 36.0 Muse Spark 1.3
  11. 35.1 Claude Haiku 5.5
  12. 35.1 Muse Wonder Full
  13. 33.3 GPT-5.6 Sol
  14. 30.6 Grok 4.6
  15. 30.2 GPT-6 Sol
  16. 27.5 Muse Wonder
  17. 26.1 Kimi K3.1 Pro
  18. 25.2 Grok 4.7
  19. 25.2 Inkling 2 Small
  20. 23.9 Qwen 3.8 Max
  21. 23.0 GLM-5.3
  22. 22.5 Ember-1
  23. 22.1 Beam 1.0
  24. 21.6 MiMo-V2.6 Flash
  25. 21.2 DeepSeek V4.1 Flash
  26. 20.7 Gemini 3.1 Pro
  27. 19.4 GPT-6 Luna
  28. 18.5 Step 5 Preview
  29. 15.3 MiMo-V2.6 Pro
  30. 14.4 GPT-5.6 Luna
  31. 12.2 Mistral Large 4 Preview
  1. 1 GPT-6 Astra 53.1%
  2. 2 Claude Fable 5.2 51.8%
  3. 3 Claude Opus 5.5 49.5%
  4. 4 Claude Sonnet 5.5 49.1%
  5. 5 Claude Fable 5.1 47.8%
  6. 6 GPT-6.1 Sol 46.4%
  7. 7 Gemini 4 Argon 45.5%
  8. 8 Claude Opus 5 37.4%
  9. 9 Gemini 3.8 Flash 36.5%
  10. 10 Muse Spark 1.3 max 36%
  11. 11 Claude Haiku 5.5 35.1%
  12. 12 Muse Wonder Full 35.1%
  13. 13 GPT-5.6 Sol 33.3%
  14. 14 Grok 4.6 30.6%
  15. 15 GPT-6 Sol 30.2%
  16. 16 Muse Wonder 27.5%
  17. 17 Kimi K3.1 Pro 26.1%
  18. 18 Grok 4.7 25.2%
  19. 19 Inkling 2 Small 25.2%
  20. 20 Qwen 3.8 Max 23.9%
  21. 21 GLM-5.3 23%
  22. 22 Ember-1 22.5%
  23. 23 Beam 1.0 22.1%
  24. 24 MiMo-V2.6 Flash 21.6%
  25. 25 DeepSeek V4.1 Flash 21.2%
  26. 26 Gemini 3.1 Pro 20.7%
  27. 27 GPT-6 Luna 19.4%
  28. 28 Step 5 Preview 18.5%
  29. 29 MiMo-V2.6 Pro 15.3%
  30. 30 GPT-5.6 Luna 14.4%
  31. 31 Mistral Large 4 Preview 12.2%

Source: Official MysteryMechanism leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 53.1% Sep 2026 Official leaderboard
2 Claude Fable 5.2 51.8% Nov 2026 Official leaderboard
3 Claude Opus 5.5 49.5% Sep 2026 Official leaderboard
4 Claude Sonnet 5.5 49.1% Sep 2026 Official leaderboard
5 Claude Fable 5.1 47.8% Sep 2026 Official leaderboard
6 GPT-6.1 Sol 46.4% Sep 2026 Official leaderboard
7 Gemini 4 Argon 45.5% Sep 2026 Official leaderboard
8 Claude Opus 5 37.4% Jul 2026 Official leaderboard
9 Gemini 3.8 Flash 36.5% Sep 2026 Official leaderboard
10 Muse Spark 1.3 36% max Sep 2026 Official leaderboard
11 Claude Haiku 5.5 35.1% Oct 2026 Official leaderboard
12 Muse Wonder Full 35.1% Oct 2026 Official leaderboard
13 GPT-5.6 Sol 33.3% Jul 2026 Official leaderboard
14 Grok 4.6 30.6% Aug 2026 Official leaderboard
15 GPT-6 Sol 30.2% Sep 2026 Official leaderboard
16 Muse Wonder 27.5% Oct 2026 Official leaderboard
17 Kimi K3.1 Pro 26.1% Oct 2026 Official leaderboard
18 Grok 4.7 25.2% Sep 2026 Official leaderboard
19 Inkling 2 Small 25.2% Nov 2026 Official leaderboard
20 Qwen 3.8 Max 23.9% Aug 2026 Official leaderboard
21 GLM-5.3 23% Aug 2026 Official leaderboard
22 Ember-1 22.5% Sep 2026 Official leaderboard
23 Beam 1.0 22.1% Oct 2026 Official leaderboard
24 MiMo-V2.6 Flash 21.6% Sep 2026 Official leaderboard
25 DeepSeek V4.1 Flash 21.2% Sep 2026 Official leaderboard
26 Gemini 3.1 Pro 20.7% Feb 2026 Official leaderboard
27 GPT-6 Luna 19.4% Sep 2026 Official leaderboard
28 Step 5 Preview 18.5% Sep 2026 Official leaderboard
29 MiMo-V2.6 Pro 15.3% Sep 2026 Official leaderboard
30 GPT-5.6 Luna 14.4% Jul 2026 Official leaderboard
31 Mistral Large 4 Preview 12.2% Oct 2026 Official leaderboard

Accuracy against latency

The Vals scatter chart with its x axis switched from Cost to Latency: mean time per test against accuracy, on a log scale. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0% 20% 40% 60% 1001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Gemini 3.1 Pro: Accuracy 20.7%, latency (mean per test) 360 s GPT-5.6 Luna: Accuracy 14.4%, latency (mean per test) 493 s GPT-5.6 Sol: Accuracy 33.3%, latency (mean per test) 523 s Claude Opus 5: Accuracy 37.4%, latency (mean per test) 1,013 s Qwen 3.8 Max: Accuracy 23.9%, latency (mean per test) 2,269 s Grok 4.6: Accuracy 30.6%, latency (mean per test) 1,675 s GLM-5.3: Accuracy 23%, latency (mean per test) 1,842 s Claude Fable 5.1: Accuracy 47.8%, latency (mean per test) 1,061 s Gemini 3.8 Flash: Accuracy 36.5%, latency (mean per test) 322 s Muse Spark 1.3 (max): Accuracy 36%, latency (mean per test) 452 s DeepSeek V4.1 Flash: Accuracy 21.2%, latency (mean per test) 484 s Step 5 Preview: Accuracy 18.5%, latency (mean per test) 1,760 s Grok 4.7: Accuracy 25.2%, latency (mean per test) 1,053 s MiMo-V2.6 Flash: Accuracy 21.6%, latency (mean per test) 2,108 s MiMo-V2.6 Pro: Accuracy 15.3%, latency (mean per test) 2,640 s Claude Opus 5.5: Accuracy 49.5%, latency (mean per test) 1,179 s GPT-6 Luna: Accuracy 19.4%, latency (mean per test) 361 s GPT-6 Sol: Accuracy 30.2%, latency (mean per test) 373 s Ember-1: Accuracy 22.5%, latency (mean per test) 1,215 s Claude Sonnet 5.5: Accuracy 49.1%, latency (mean per test) 1,410 s GPT-6.1 Sol: Accuracy 46.4%, latency (mean per test) 437 s Gemini 4 Argon: Accuracy 45.5%, latency (mean per test) 1,495 s Mistral Large 4 Preview: Accuracy 12.2%, latency (mean per test) 2,137 s Claude Haiku 5.5: Accuracy 35.1%, latency (mean per test) 805 s Muse Wonder: Accuracy 27.5%, latency (mean per test) 1,097 s Muse Wonder Full: Accuracy 35.1%, latency (mean per test) 2,494 s Kimi K3.1 Pro: Accuracy 26.1%, latency (mean per test) 360 s Inkling 2 Small: Accuracy 25.2%, latency (mean per test) 1,017 s GPT-6 Astra: Accuracy 53.1%, latency (mean per test) 489 s Beam 1.0: Accuracy 22.1%, latency (mean per test) 144 s Claude Fable 5.2: Accuracy 51.8%, latency (mean per test) 292 s GPT-6 AstraClaude Fable 5.2Beam 1.0 0% 20% 40% 60% 1001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Gemini 3.1 Pro: Accuracy 20.7%, latency (mean per test) 360 s GPT-5.6 Luna: Accuracy 14.4%, latency (mean per test) 493 s GPT-5.6 Sol: Accuracy 33.3%, latency (mean per test) 523 s Claude Opus 5: Accuracy 37.4%, latency (mean per test) 1,013 s Qwen 3.8 Max: Accuracy 23.9%, latency (mean per test) 2,269 s Grok 4.6: Accuracy 30.6%, latency (mean per test) 1,675 s GLM-5.3: Accuracy 23%, latency (mean per test) 1,842 s Claude Fable 5.1: Accuracy 47.8%, latency (mean per test) 1,061 s Gemini 3.8 Flash: Accuracy 36.5%, latency (mean per test) 322 s Muse Spark 1.3 (max): Accuracy 36%, latency (mean per test) 452 s DeepSeek V4.1 Flash: Accuracy 21.2%, latency (mean per test) 484 s Step 5 Preview: Accuracy 18.5%, latency (mean per test) 1,760 s Grok 4.7: Accuracy 25.2%, latency (mean per test) 1,053 s MiMo-V2.6 Flash: Accuracy 21.6%, latency (mean per test) 2,108 s MiMo-V2.6 Pro: Accuracy 15.3%, latency (mean per test) 2,640 s Claude Opus 5.5: Accuracy 49.5%, latency (mean per test) 1,179 s GPT-6 Luna: Accuracy 19.4%, latency (mean per test) 361 s GPT-6 Sol: Accuracy 30.2%, latency (mean per test) 373 s Ember-1: Accuracy 22.5%, latency (mean per test) 1,215 s Claude Sonnet 5.5: Accuracy 49.1%, latency (mean per test) 1,410 s GPT-6.1 Sol: Accuracy 46.4%, latency (mean per test) 437 s Gemini 4 Argon: Accuracy 45.5%, latency (mean per test) 1,495 s Mistral Large 4 Preview: Accuracy 12.2%, latency (mean per test) 2,137 s Claude Haiku 5.5: Accuracy 35.1%, latency (mean per test) 805 s Muse Wonder: Accuracy 27.5%, latency (mean per test) 1,097 s Muse Wonder Full: Accuracy 35.1%, latency (mean per test) 2,494 s Kimi K3.1 Pro: Accuracy 26.1%, latency (mean per test) 360 s Inkling 2 Small: Accuracy 25.2%, latency (mean per test) 1,017 s GPT-6 Astra: Accuracy 53.1%, latency (mean per test) 489 s Beam 1.0: Accuracy 22.1%, latency (mean per test) 144 s Claude Fable 5.2: Accuracy 51.8%, latency (mean per test) 292 s GPT-6 AstraClaude Fable 5.2Beam 1.0
ModelSettingAccuracyLatency (mean per test)
GPT-6 Astra 53.2% 488.6 s
Claude Fable 5.2 51.8% 292.06 s
Claude Opus 5.5 49.6% 1,179 s
Claude Sonnet 5.5 49.1% 1,410 s
Claude Fable 5.1 47.8% 1,061 s
GPT-6.1 Sol 46.4% 437.02 s
Gemini 4 Argon 45.5% 1,495 s
Claude Opus 5 37.4% 1,013 s
Gemini 3.8 Flash 36.5% 322.24 s
Muse Spark 1.3 max 36% 451.8 s
Claude Haiku 5.5 35.1% 805.03 s
Muse Wonder Full 35.1% 2,494 s
GPT-5.6 Sol 33.3% 522.74 s
Grok 4.6 30.6% 1,675 s
GPT-6 Sol 30.2% 373.43 s
Muse Wonder 27.5% 1,097 s
Kimi K3.1 Pro 26.1% 359.9 s
Inkling 2 Small 25.2% 1,017 s
Grok 4.7 25.2% 1,053 s
Qwen 3.8 Max 23.9% 2,269 s
GLM-5.3 23% 1,842 s
Ember-1 22.5% 1,215 s
Beam 1.0 22.1% 144.31 s
MiMo-V2.6 Flash 21.6% 2,108 s
DeepSeek V4.1 Flash 21.2% 484.33 s
Gemini 3.1 Pro 20.7% 360.46 s
GPT-6 Luna 19.4% 361.21 s
Step 5 Preview 18.5% 1,760 s
MiMo-V2.6 Pro 15.3% 2,640 s
GPT-5.6 Luna 14.4% 493.05 s
Mistral Large 4 Preview 12.2% 2,137 s

The best score, 53.1%, costs $1.56 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.

OpenAI Xiaomi Corp Thinking Machines Anthropic Cost frontier One model at each effort level
0% 20% 40% 60% $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra GPT-6.1 Sol Inkling 2 Small MiMo-V2.6 Flash GPT-6 Luna 0% 20% 40% 60% $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-6 Astra GPT-6.1 Sol

Source: Official MysteryMechanism leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra 53.1% $1.56 21,201
Claude Fable 5.2 51.8% $1.55 27,779
Claude Opus 5.5 49.5% $4.06 131,452
Claude Sonnet 5.5 49.1% $2.98 193,418
Claude Fable 5.1 47.8% $5.63 80,806
GPT-6.1 Sol 46.4% $0.28 20,186
Gemini 4 Argon 45.5% $4.59 96,886
Claude Opus 5 37.4% $3.28 81,310
Gemini 3.8 Flash 36.5% $1.73 80,442
Muse Spark 1.3 max 36% $0.85 91,046
Claude Haiku 5.5 35.1% $0.54 186,893
Muse Wonder Full 35.1% $4.16 99,500
GPT-5.6 Sol 33.3% $0.96 27,506
Grok 4.6 30.6% $1.31 107,454
GPT-6 Sol 30.2% $0.46 30,502
Muse Wonder 27.5% $3.39 102,039
Kimi K3.1 Pro 26.1% $0.62 38,095
Grok 4.7 25.2% $1.84 80,598
Inkling 2 Small 25.2% $0.16 84,377
Qwen 3.8 Max 23.9% $1.31 108,723
GLM-5.3 23% $1.53 156,748
Ember-1 22.5% $1.29 60,449
Beam 1.0 22.1% 36,624
MiMo-V2.6 Flash 21.6% $0.06 115,716
DeepSeek V4.1 Flash 21.2% $0.16
Gemini 3.1 Pro 20.7% $1.07 36,840
GPT-6 Luna 19.4% $0.03 36,248
Step 5 Preview 18.5% $0.60 126,039
MiMo-V2.6 Pro 15.3% $0.15 111,531
GPT-5.6 Luna 14.4% $0.11 42,721
Mistral Large 4 Preview 12.2% $1.06 236,882

2 models have held number one on the official leaderboard since its release in September 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic OpenAI
Own leaderboard 2 leaders
GPT-6 Astra 53.1%

Source: Official MysteryMechanism leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra Claude Fable 5.1 Claude Opus 5 Gemini 3.1 Pro 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra Claude Opus 5

Source: Official MysteryMechanism leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Science and reasoning

All 92 in Science and reasoning
  • GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics. Gemini 4 Carbon 96.5%
  • Terminal-Bench-Science In use 70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate. GPT-6 Astra 68.1%
  • FrontierScience In use Models solve physics, chemistry, and biology problems in an Olympiad track graded by short answer and a Research track graded by 10-point rubrics. Intern-S2-397B 87%
  • BioMysteryBench In use Bioinformatics problems on anonymized real datasets that require real analysis with tools. Rubric-graded, with human-solvable and human-difficult splits. Claude Fable 5.2 83.7%
  • HiPhO In use Models solve theory problems from 2024 and 2025 physics olympiad exams, some with diagrams. Official marking schemes grade answers and steps, mapped to medal levels. Gemini 3 Pro 78.3%
  • LABBench2 In use 11 biology research sub-tasks (literature, figures, tables, patents, trials and more) with tool access. Macro-averaged accuracy. GPT-5.2 60.9%