Benchmark Almanac

Science's Last Exam

GPT-6.1 Sol 44.7 leads since November 2026
Not in model cards

Agents conduct scientific research through tools, code, desktop software, laboratory recordings, and withheld references across 113 tasks in nine disciplines. The official composite equally averages six assessment scores on a zero-to-100 scale. Checks combine reference comparisons, executable tests, and literature audits.

AItonomy requires separate approval for hidden reference data. The composite includes the current zero scores for end-to-end research but excludes wet-lab execution, which has no released tasks. The board averages cost and time over four capabilities only. Models lacking an assessment receive no composite.

Released
November 2026
Kind
Research science
Level
Research frontier
Scored
Checked automatically
Answers
Agent task
Tasks
Public, plus a private test set
Strongest model tested
GPT-6.1 Sol #5 on the capability index
Last model added
October 2026
Scores refreshed
November 12th, 2026
9
Models scored on its own leaderboard

GPT-6.1 Sol leads with 44.7

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.

  1. 44.7 GPT-6.1 Sol
  2. 37.5 Claude Opus 5
  3. 32.7 Gemini 4 Argon
  4. 28.9 Kimi K3.1 Pro
  5. 26.8 DeepSeek V4.1 Flash
  6. 26.4 MiMo-V2.6 Pro
  7. 23.3 Qwen 3.8 Max
  8. 21.3 GLM-5.3-Flash
  9. 16.8 MiniMax-M3
  1. 1 GPT-6.1 Sol high 44.7
  2. 2 Claude Opus 5 high 37.5
  3. 3 Gemini 4 Argon default 32.7
  4. 4 Kimi K3.1 Pro high 28.9
  5. 5 DeepSeek V4.1 Flash high 26.8
  6. 6 MiMo-V2.6 Pro high 26.4
  7. 7 Qwen 3.8 Max default 23.3
  8. 8 GLM-5.3-Flash high 21.3
  9. 9 MiniMax-M3 high 16.8

Source: Official Science's Last Exam leaderboard

#ModelScoreSettingDateSource
1 GPT-6.1 Sol 44.7 high, Codex; DiscoveryRSI for Hypothesis Discovery Sep 2026 Official leaderboard
2 Claude Opus 5 37.5 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Jul 2026 Official leaderboard
3 Gemini 4 Argon 32.7 default, Codex; DiscoveryRSI for Hypothesis Discovery Sep 2026 Official leaderboard
4 Kimi K3.1 Pro 28.9 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Oct 2026 Official leaderboard
5 DeepSeek V4.1 Flash 26.8 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Sep 2026 Official leaderboard
6 MiMo-V2.6 Pro 26.4 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Sep 2026 Official leaderboard
7 Qwen 3.8 Max 23.3 default, Claude Code; DiscoveryRSI for Hypothesis Discovery Aug 2026 Official leaderboard
8 GLM-5.3-Flash 21.3 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Aug 2026 Official leaderboard
9 MiniMax-M3 16.8 high, Claude Code; DiscoveryRSI for Hypothesis Discovery Jun 2026 Official leaderboard

Capability vs. cost

SLE averages cost across four capabilities while its composite equally weights six assessments. 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 10 20 30 40 50 $0$5$10$15$20 Cost per task ($), lower is better Composite score (score) MiniMax-M3 (high): Composite score 16.8 score, cost per task $1.15 Claude Opus 5 (high): Composite score 37.5 score, cost per task $19.7 Qwen 3.8 Max (default): Composite score 23.3 score, cost per task $6.43 MiMo-V2.6 Pro (high): Composite score 26.4 score, cost per task $0.56 Gemini 4 Argon (default): Composite score 32.7 score, cost per task $4.26 GLM-5.3-Flash (high): Composite score 21.3 score, cost per task $0.38 DeepSeek V4.1 Flash (high): Composite score 26.8 score, cost per task $0.42 GPT-6.1 Sol (high): Composite score 44.7 score, cost per task $3.07 Kimi K3.1 Pro (high): Composite score 28.9 score, cost per task $2.89 GPT-6.1 Sol (high)Kimi K3.1 Pro (high)DeepSeek V4.1 Flash (high)GLM-5.3-Flash (high) 0 10 20 30 40 50 $0$5$10$15$20 Cost per task ($), lower is better Composite score (score) MiniMax-M3 (high): Composite score 16.8 score, cost per task $1.15 Claude Opus 5 (high): Composite score 37.5 score, cost per task $19.7 Qwen 3.8 Max (default): Composite score 23.3 score, cost per task $6.43 MiMo-V2.6 Pro (high): Composite score 26.4 score, cost per task $0.56 Gemini 4 Argon (default): Composite score 32.7 score, cost per task $4.26 GLM-5.3-Flash (high): Composite score 21.3 score, cost per task $0.38 DeepSeek V4.1 Flash (high): Composite score 26.8 score, cost per task $0.42 GPT-6.1 Sol (high): Composite score 44.7 score, cost per task $3.07 Kimi K3.1 Pro (high): Composite score 28.9 score, cost per task $2.89 GPT-6.1 Sol (high)Kimi K3.1 Pro (high)DeepSeek V4.1 Flash (high)GLM-5.3-Flash (high)
ModelSettingComposite scoreCost per task
GPT-6.1 Sol high, Codex; DiscoveryRSI for Hypothesis Discovery 44.66 score $3.07
Claude Opus 5 high, Claude Code; DiscoveryRSI for Hypothesis Discovery 37.5 score $19.7
Gemini 4 Argon default, Codex; DiscoveryRSI for Hypothesis Discovery 32.74 score $4.26
Kimi K3.1 Pro high, Claude Code; DiscoveryRSI for Hypothesis Discovery 28.92 score $2.89
DeepSeek V4.1 Flash high, Claude Code; DiscoveryRSI for Hypothesis Discovery 26.83 score $0.42
MiMo-V2.6 Pro high, Claude Code; DiscoveryRSI for Hypothesis Discovery 26.4 score $0.56
Qwen 3.8 Max default, Claude Code; DiscoveryRSI for Hypothesis Discovery 23.3 score $6.43
GLM-5.3-Flash high, Claude Code; DiscoveryRSI for Hypothesis Discovery 21.33 score $0.38
MiniMax-M3 high, Claude Code; DiscoveryRSI for Hypothesis Discovery 16.82 score $1.15

The best score, 44.7, costs $3.07 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.

Z.ai DeepSeek OpenAI Moonshot AI Cost frontier One model at each effort level
0.00 10.0 20.0 30.0 40.0 50.0 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6.1 Sol, high Kimi K3.1 Pro, high DeepSeek V4.1 Flash, high GLM-5.3-Flash, high 0.00 10.0 20.0 30.0 40.0 50.0 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6.1 Sol, high Kimi K3.1 Pro, high

Source: Official Science's Last Exam leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6.1 Sol high Codex; DiscoveryRSI for Hypothesis Discovery 44.7 $3.07
Claude Opus 5 high Claude Code; DiscoveryRSI for Hypothesis Discovery 37.5 $20
Gemini 4 Argon default Codex; DiscoveryRSI for Hypothesis Discovery 32.7 $4.26
Kimi K3.1 Pro high Claude Code; DiscoveryRSI for Hypothesis Discovery 28.9 $2.89
DeepSeek V4.1 Flash high Claude Code; DiscoveryRSI for Hypothesis Discovery 26.8 $0.42
MiMo-V2.6 Pro high Claude Code; DiscoveryRSI for Hypothesis Discovery 26.4 $0.56
Qwen 3.8 Max default Claude Code; DiscoveryRSI for Hypothesis Discovery 23.3 $6.43
GLM-5.3-Flash high Claude Code; DiscoveryRSI for Hypothesis Discovery 21.3 $0.38
MiniMax-M3 high Claude Code; DiscoveryRSI for Hypothesis Discovery 16.8 $1.15

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-6.1 Sol 44.7

Source: Official Science's Last Exam 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
10.0 20.0 30.0 40.0 50.0 2026 2027 10.0 20.0 30.0 40.0 50.0 2026 2027

Source: Official Science's Last Exam 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%