Inference
EN
Vals AI

MysteryMechanism

Can agents rediscover sealed mathematical mechanisms through bounded experiments?

ModelsModels with a result.30
Top resultBest result on this benchmark.53.2%GPT-6 Astra
Top-3 spreadPoints from first to third.3.6 pts
UpdatedDate the source changed the results.11 Nov 2026

Result and cost

29 of 30 results

Results

30 results
#ModelResultThe score from the source.CostUS dollars to run one task.TimeTime to run one task.
01GPT-6 Astra53.2% ±3.4$1.568 min
02Claude Fable 5.251.8% ±3.4$1.555 min
03Claude Opus 5.549.5% ±3.4$4.0620 min
04Claude Sonnet 5.549.1% ±3.4$2.9824 min
05Claude Fable 5.147.7% ±3.4$5.6318 min
06GPT-6.1 Sol46.4% ±3.4$0.2857 min
07Gemini 4 Argon45.5% ±3.4$4.5925 min
08Claude Opus 537.4% ±3.3$3.2817 min
09Gemini 3.8 Flash36.5% ±3.2$1.735 min
10Muse Spark 1.3 Max36.0% ±3.2$0.8528 min
11Muse Wonder Full35.1% ±3.2$4.1642 min
11Claude Haiku 5.535.1% ±3.2$0.54113 min
13GPT-5.6 Sol33.3% ±3.2$0.9639 min
14Grok 4.630.6% ±3.1$1.3128 min
15GPT-6 Sol30.2% ±3.1$0.4616 min
16Muse Wonder27.5% ±3.0$3.3918 min
17Grok 4.725.2% ±2.9$1.8418 min
17Inkling 2 Small25.2% ±2.9$0.1617 min
19Qwen3.8 Max23.9% ±2.9$1.3138 min
20GLM-5.323.0% ±2.8$1.5331 min
21Ember-122.5% ±2.8$1.2920 min
22Beam 1.022.1% ±2.8—2 min
23MiMo-V2.6-Flash21.6% ±2.8$0.05735 min
24DeepSeek V4.1 Flash 101721.2% ±2.7$0.1648 min
25Gemini 3.1 Pro Preview20.7% ±2.7$1.076 min
26GPT-6 Luna19.4% ±2.7$0.0266 min
27Step 5 Preview18.5% ±2.6$0.59729 min
28MiMo-V2.6-Pro15.3% ±2.4$0.15444 min
29GPT-5.6 Luna14.4% ±2.4$0.1098 min
30Mistral Large 412.2% ±2.2$1.0636 min