Benchmark Almanac

Vals Science Index

GPT-6 Astra 65.1% leads since October 2026
Not in model cards

Agents are scored on the equal average of BioMysteryBench, MysteryMechanism and Terminal-Bench-Science.

Version 1.0 released October 16th, 2026. The board had 23 models on October 25th, 2026.

Made by
Vals AI
Released
October 2026
Kind
Composite indices
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
25
Models scored on its own leaderboard

GPT-6 Astra leads with 65.1%

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.

25 models. Scroll sideways for the rest.

  1. 65.1 GPT-6 Astra
  2. 63.7 Claude Fable 5.2
  3. 59.6 GPT-6.1 Sol
  4. 59.1 Claude Sonnet 5.5
  5. 58.6 Claude Opus 5.5
  6. 55.4 Gemini 4 Argon
  7. 48.5 Muse Wonder Full
  8. 47.9 Claude Opus 5
  9. 45.0 GPT-6 Sol
  10. 44.4 Kimi K3.1 Pro
  11. 41.5 GPT-5.6 Sol
  12. 39.1 Muse Wonder
  13. 38.5 Claude Haiku 5.5
  14. 38.1 Grok 4.6
  15. 35.8 Gemini 3.8 Flash
  16. 34.8 Grok 4.7
  17. 32.7 Inkling 2 Small
  18. 31.8 Ember-1
  19. 31.7 MiMo-V2.6 Flash
  20. 29.6 DeepSeek V4.1 Flash
  21. 29.6 Step 5 Preview
  22. 28.4 GPT-6 Luna
  23. 27.4 Gemini 3.1 Pro
  24. 25.3 GPT-5.6 Luna
  25. 22.0 Beam 1.0
  1. 1 GPT-6 Astra 65.1%
  2. 2 Claude Fable 5.2 63.7%
  3. 3 GPT-6.1 Sol 59.6%
  4. 4 Claude Sonnet 5.5 59.1%
  5. 5 Claude Opus 5.5 58.6%
  6. 6 Gemini 4 Argon 55.4%
  7. 7 Muse Wonder Full 48.5%
  8. 8 Claude Opus 5 47.9%
  9. 9 GPT-6 Sol 45%
  10. 10 Kimi K3.1 Pro 44.4%
  11. 11 GPT-5.6 Sol 41.5%
  12. 12 Muse Wonder 39.1%
  13. 13 Claude Haiku 5.5 38.5%
  14. 14 Grok 4.6 38.1%
  15. 15 Gemini 3.8 Flash 35.8%
  16. 16 Grok 4.7 34.8%
  17. 17 Inkling 2 Small 32.7%
  18. 18 Ember-1 31.8%
  19. 19 MiMo-V2.6 Flash 31.7%
  20. 20 DeepSeek V4.1 Flash 29.6%
  21. 21 Step 5 Preview 29.6%
  22. 22 GPT-6 Luna 28.4%
  23. 23 Gemini 3.1 Pro 27.4%
  24. 24 GPT-5.6 Luna 25.3%
  25. 25 Beam 1.0 22%

Source: Official Vals Science Index leaderboard

#ModelScoreSettingDateSource
1 GPT-6 Astra 65.1% Sep 2026 Official leaderboard
2 Claude Fable 5.2 63.7% Nov 2026 Official leaderboard
3 GPT-6.1 Sol 59.6% Sep 2026 Official leaderboard
4 Claude Sonnet 5.5 59.1% Sep 2026 Official leaderboard
5 Claude Opus 5.5 58.6% Sep 2026 Official leaderboard
6 Gemini 4 Argon 55.4% Sep 2026 Official leaderboard
7 Muse Wonder Full 48.5% Oct 2026 Official leaderboard
8 Claude Opus 5 47.9% Jul 2026 Official leaderboard
9 GPT-6 Sol 45% Sep 2026 Official leaderboard
10 Kimi K3.1 Pro 44.4% Oct 2026 Official leaderboard
11 GPT-5.6 Sol 41.5% Jul 2026 Official leaderboard
12 Muse Wonder 39.1% Oct 2026 Official leaderboard
13 Claude Haiku 5.5 38.5% Oct 2026 Official leaderboard
14 Grok 4.6 38.1% Aug 2026 Official leaderboard
15 Gemini 3.8 Flash 35.8% Sep 2026 Official leaderboard
16 Grok 4.7 34.8% Sep 2026 Official leaderboard
17 Inkling 2 Small 32.7% Nov 2026 Official leaderboard
18 Ember-1 31.8% Sep 2026 Official leaderboard
19 MiMo-V2.6 Flash 31.7% Sep 2026 Official leaderboard
20 DeepSeek V4.1 Flash 29.6% Sep 2026 Official leaderboard
21 Step 5 Preview 29.6% Sep 2026 Official leaderboard
22 GPT-6 Luna 28.4% Sep 2026 Official leaderboard
23 Gemini 3.1 Pro 27.4% Feb 2026 Official leaderboard
24 GPT-5.6 Luna 25.3% Jul 2026 Official leaderboard
25 Beam 1.0 22% 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% 80% 1,0002,0005,00010K Latency (mean per test) (s), log scale, lower is better Accuracy GPT-5.6 Luna: Accuracy 25.3%, latency (mean per test) 3,000 s GPT-5.6 Sol: Accuracy 41.5%, latency (mean per test) 2,612 s Claude Opus 5: Accuracy 47.9%, latency (mean per test) 3,706 s Grok 4.6: Accuracy 38.1%, latency (mean per test) 2,410 s DeepSeek V4.1 Flash: Accuracy 29.6%, latency (mean per test) 3,999 s Step 5 Preview: Accuracy 29.6%, latency (mean per test) 4,018 s Grok 4.7: Accuracy 34.8%, latency (mean per test) 2,434 s MiMo-V2.6 Flash: Accuracy 31.7%, latency (mean per test) 6,048 s Claude Opus 5.5: Accuracy 58.6%, latency (mean per test) 3,591 s GPT-6 Luna: Accuracy 28.4%, latency (mean per test) 1,815 s Ember-1: Accuracy 31.8%, latency (mean per test) 4,659 s Claude Sonnet 5.5: Accuracy 59.1%, latency (mean per test) 4,159 s Gemini 4 Argon: Accuracy 55.4%, latency (mean per test) 3,420 s Beam 1.0: Accuracy 22%, latency (mean per test) 1,748 s Claude Haiku 5.5: Accuracy 38.5%, latency (mean per test) 2,398 s Muse Wonder: Accuracy 39.1%, latency (mean per test) 2,588 s Muse Wonder Full: Accuracy 48.5%, latency (mean per test) 5,663 s Inkling 2 Small: Accuracy 32.7%, latency (mean per test) 3,449 s Gemini 3.1 Pro: Accuracy 27.4%, latency (mean per test) 1,084 s Gemini 3.8 Flash: Accuracy 35.8%, latency (mean per test) 1,320 s GPT-6 Astra: Accuracy 65.1%, latency (mean per test) 2,437 s GPT-6 Sol: Accuracy 45%, latency (mean per test) 1,876 s GPT-6.1 Sol: Accuracy 59.6%, latency (mean per test) 1,991 s Kimi K3.1 Pro: Accuracy 44.4%, latency (mean per test) 1,382 s Claude Fable 5.2: Accuracy 63.7%, latency (mean per test) 2,081 s GPT-6 AstraClaude Fable 5.2GPT-6.1 SolGPT-6 SolKimi K3.1 ProGemini 3.8 FlashGemini 3.1 Pro 0% 20% 40% 60% 80% 1,0002,0005,00010K Latency (mean per test) (s), log scale, lower is better Accuracy GPT-5.6 Luna: Accuracy 25.3%, latency (mean per test) 3,000 s GPT-5.6 Sol: Accuracy 41.5%, latency (mean per test) 2,612 s Claude Opus 5: Accuracy 47.9%, latency (mean per test) 3,706 s Grok 4.6: Accuracy 38.1%, latency (mean per test) 2,410 s DeepSeek V4.1 Flash: Accuracy 29.6%, latency (mean per test) 3,999 s Step 5 Preview: Accuracy 29.6%, latency (mean per test) 4,018 s Grok 4.7: Accuracy 34.8%, latency (mean per test) 2,434 s MiMo-V2.6 Flash: Accuracy 31.7%, latency (mean per test) 6,048 s Claude Opus 5.5: Accuracy 58.6%, latency (mean per test) 3,591 s GPT-6 Luna: Accuracy 28.4%, latency (mean per test) 1,815 s Ember-1: Accuracy 31.8%, latency (mean per test) 4,659 s Claude Sonnet 5.5: Accuracy 59.1%, latency (mean per test) 4,159 s Gemini 4 Argon: Accuracy 55.4%, latency (mean per test) 3,420 s Beam 1.0: Accuracy 22%, latency (mean per test) 1,748 s Claude Haiku 5.5: Accuracy 38.5%, latency (mean per test) 2,398 s Muse Wonder: Accuracy 39.1%, latency (mean per test) 2,588 s Muse Wonder Full: Accuracy 48.5%, latency (mean per test) 5,663 s Inkling 2 Small: Accuracy 32.7%, latency (mean per test) 3,449 s Gemini 3.1 Pro: Accuracy 27.4%, latency (mean per test) 1,084 s Gemini 3.8 Flash: Accuracy 35.8%, latency (mean per test) 1,320 s GPT-6 Astra: Accuracy 65.1%, latency (mean per test) 2,437 s GPT-6 Sol: Accuracy 45%, latency (mean per test) 1,876 s GPT-6.1 Sol: Accuracy 59.6%, latency (mean per test) 1,991 s Kimi K3.1 Pro: Accuracy 44.4%, latency (mean per test) 1,382 s Claude Fable 5.2: Accuracy 63.7%, latency (mean per test) 2,081 s GPT-6 AstraClaude Fable 5.2GPT-6.1 SolGPT-6 SolKimi K3.1 ProGemini 3.8 FlashGemini 3.1 Pro
ModelSettingAccuracyLatency (mean per test)
GPT-6 Astra 65.1% 2,437 s
Claude Fable 5.2 63.7% 2,081 s
GPT-6.1 Sol 59.6% 1,991 s
Claude Sonnet 5.5 59.1% 4,159 s
Claude Opus 5.5 58.7% 3,591 s
Gemini 4 Argon 55.4% 3,420 s
Muse Wonder Full 48.5% 5,663 s
Claude Opus 5 47.9% 3,706 s
GPT-6 Sol 45% 1,876 s
Kimi K3.1 Pro 44.4% 1,382 s
GPT-5.6 Sol 41.5% 2,612 s
Muse Wonder 39.1% 2,588 s
Claude Haiku 5.5 38.5% 2,398 s
Grok 4.6 38.1% 2,410 s
Gemini 3.8 Flash 35.8% 1,320 s
Grok 4.7 34.8% 2,434 s
Inkling 2 Small 32.7% 3,449 s
Ember-1 31.8% 4,659 s
MiMo-V2.6 Flash 31.7% 6,048 s
DeepSeek V4.1 Flash 29.7% 3,999 s
Step 5 Preview 29.6% 4,018 s
GPT-6 Luna 28.4% 1,815 s
Gemini 3.1 Pro 27.4% 1,084 s
GPT-5.6 Luna 25.3% 3,000 s
Beam 1.0 22% 1,748 s

The best score, 65.1%, costs $8.03 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 Cost frontier One model at each effort level
0% 20% 40% 60% 80% $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra GPT-6.1 Sol MiMo-V2.6 Flash GPT-6 Luna 0% 20% 40% 60% 80% $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra GPT-6.1 Sol

Source: Official Vals Science Index leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra 65.1% $8.03 65,321
Claude Fable 5.2 63.7% $8.79 94,053
GPT-6.1 Sol 59.6% $1.36 68,765
Claude Sonnet 5.5 59.1% $8.50 281,131
Claude Opus 5.5 58.6% $8.71 150,133
Gemini 4 Argon 55.4% $18 196,916
Muse Wonder Full 48.5% $10 136,611
Claude Opus 5 47.9% $13
GPT-6 Sol 45% $2.31 61,916
Kimi K3.1 Pro 44.4% $3.33 71,899
GPT-5.6 Sol 41.5% $3.40
Muse Wonder 39.1% $8.74 137,061
Claude Haiku 5.5 38.5% $1.86 248,672
Grok 4.6 38.1% $2.45 85,489
Gemini 3.8 Flash 35.8% $2.98 94,594
Grok 4.7 34.8% $7.12 97,507
Inkling 2 Small 32.7% $1.99 208,107
Ember-1 31.8% $11 122,076
MiMo-V2.6 Flash 31.7% $0.12 157,496
DeepSeek V4.1 Flash 29.6% $0.29
Step 5 Preview 29.6% $1.78 170,536
GPT-6 Luna 28.4% $0.11 76,098
Gemini 3.1 Pro 27.4% $2.56 39,111
GPT-5.6 Luna 25.3% $0.26
Beam 1.0 22% 90,658

Number one over time

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

OpenAI
Own leaderboard 1 leader
GPT-6 Astra 65.1%

Source: Official Vals Science Index 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 Opus 5 GPT-5.6 Sol Gemini 3.1 Pro 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra Claude Opus 5

Source: Official Vals Science Index 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 Other

All 43 in Other
  • Personal Superintelligence Index (Meta) In use Meta's composite of seven personal and knowledge-work evals: Agents' Last Exam, AutomationBench-AA, FocusBench, JobBench, GDP.pdf, VBT and DRACO.
  • Artificial Analysis Intelligence Index v5 In use Version 5 of Artificial Analysis's index: nine evaluations in four equal categories, most with private test sets, run three times each in its own agent harness. Gemini 4 Carbon 59.0
  • SAGE v1.1 (Vals) In use Models grade handwritten undergraduate math and computer science work against rubrics. Scored by class-balanced accuracy of each rubric check against expert grading. Claude Opus 5.5 91.3%
  • MTEB Fading Massive Text Embedding Benchmark: retrieval, clustering, classification across languages and modalities. sonar 42%
  • Anthropic ECI (AECI) Only in system cards Index that combines a basket of capability benchmarks into one capability score to track trends across Anthropic models. Claude Opus 5.5 169.4
  • SPEED-Bench Only in system cards Measures speculative decoding across varied prompts and domains. Reports the average acceptance length of draft tokens, so higher means faster generation at the same output. Nemotron 3 Ultra 5.15