Benchmark Almanac

Terminal-Bench-Science

GPT-6 Astra 68.1% leads since Sep 2026 In use

70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate.

70 tasks kept from 920 proposals. v0.1. V0.2 in progress. Citation is a Zenodo DOI, no paper.

Made by Stanford-led community (Terminal-Bench team + domain experts) Released Aug 2026 Research science All tasks public Last in a launch post Oct 2026 Last model added Oct 2026 Scores refreshed 2 Nov 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
harbor-framework/terminal-bench-science 798 stars Dataset Leaderboard
17
Models scored on its own leaderboard
8
Launch posts list it, from 4 labs
5
System cards report it, from 2 labs

GPT-6 Astra leads with 68.1%

Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

17 models. Scroll sideways for the rest.

  1. 68.1 GPT-6 Astra
  2. 63.3 Claude Opus 5.5
  3. 40.0 Claude Fable 5.1
  4. 30.0 Claude Opus 5
  5. 22.4 GPT-5.6 Sol
  6. 21.4 Claude Fable 5
  7. 15.7 DeepSeek V4.1 Flash
  8. 14.3 Grok 4.7
  9. 12.4 Gemini 3.8 Flash
  10. 10.5 Claude Opus 4.8
  11. 8.57 GPT-5.6 Terra
  12. 8.10 GLM-5.3
  13. 7.14 Kimi K3
  14. 7.14 Grok 4.6
  15. 5.71 Gemini 3.7 Flash
  16. 3.81 DeepSeek-V4-Pro
  17. 3.33 GPT-5.6 Luna
  1. 1 GPT-6 Astra max 68.1%
  2. 2 Claude Opus 5.5 max 63.3%
  3. 3 Claude Fable 5.1 max 40%
  4. 4 Claude Opus 5 max 30%
  5. 5 GPT-5.6 Sol max 22.4%
  6. 6 Claude Fable 5 max 21.4%
  7. 7 DeepSeek V4.1 Flash max 15.7%
  8. 8 Grok 4.7 xhigh 14.3%
  9. 9 Gemini 3.8 Flash high 12.4%
  10. 10 Claude Opus 4.8 max 10.5%
  11. 11 GPT-5.6 Terra max 8.57%
  12. 12 GLM-5.3 max 8.10%
  13. 13 Kimi K3 max 7.14%
  14. 14 Grok 4.6 xhigh 7.14%
  15. 15 Gemini 3.7 Flash high 5.71%
  16. 16 DeepSeek-V4-Pro max 3.81%
  17. 17 GPT-5.6 Luna max 3.33%
#ModelScoreSettingDateSource
1 GPT-6 Astra 68.1% max, Codex Sep 2026 Official leaderboard
2 Claude Opus 5.5 63.3% max, Claude Code Sep 2026 Official leaderboard
3 Claude Fable 5.1 40% max, Claude Code Sep 2026 Official leaderboard
4 Claude Opus 5 30% max, Claude Code Jul 2026 Official leaderboard
5 GPT-5.6 Sol 22.4% max, Codex Jul 2026 Official leaderboard
6 Claude Fable 5 21.4% max, Claude Code Jun 2026 Official leaderboard
7 DeepSeek V4.1 Flash 15.7% max, Codex Sep 2026 Official leaderboard
8 Grok 4.7 14.3% xhigh, Grok Build Sep 2026 Official leaderboard
9 Gemini 3.8 Flash 12.4% high, mini-SWE-agent Sep 2026 Official leaderboard
10 Claude Opus 4.8 10.5% max, Claude Code May 2026 Official leaderboard
11 GPT-5.6 Terra 8.57% max, Codex Jul 2026 Official leaderboard
12 GLM-5.3 8.10% max, Claude Code Aug 2026 Official leaderboard
13 Kimi K3 7.14% max, Claude Code Jul 2026 Official leaderboard
14 Grok 4.6 7.14% xhigh, Grok Build Aug 2026 Official leaderboard
15 Gemini 3.7 Flash 5.71% high, mini-SWE-agent Aug 2026 Official leaderboard
16 DeepSeek-V4-Pro 3.81% max, Codex Apr 2026 Official leaderboard
17 GPT-5.6 Luna 3.33% max, Codex Jul 2026 Official leaderboard

Labs report scores for 9 models in their cards. We plot every source in the charts below.

The best score, 68.1%, costs $24 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 DeepSeek Anthropic Cost frontier One model at each effort level
0% 20% 40% 60% 80% $1.00 $2.00 $5.00 $10 $20 $50 $100 Cost per task, log scale GPT-6 Astra, max Claude Opus 5.5, max GPT-5.6 Sol, max DeepSeek V4.1 Flash, max GPT-5.6 Luna, max 0% 20% 40% 60% 80% $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, max Claude Opus 5.5, max
ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra max Codex 68.1% $24
Claude Opus 5.5 max Claude Code 63.3% $23
Claude Fable 5.1 max Claude Code 40% $30
Claude Opus 5 max Claude Code 30% $33
GPT-5.6 Sol max Codex 22.4% $20
Claude Fable 5 max Claude Code 21.4% $68
DeepSeek V4.1 Flash max Codex 15.7% $1.84
Grok 4.7 xhigh Grok Build 14.3% $15
Gemini 3.8 Flash high mini-SWE-agent 12.4% $5.32
Claude Opus 4.8 max Claude Code 10.5% $28
GPT-5.6 Terra max Codex 8.57% $9.45
GLM-5.3 max Claude Code 8.10% $13
Kimi K3 max Claude Code 7.14% $7.26
Grok 4.6 xhigh Grok Build 7.14% $16
Gemini 3.7 Flash high mini-SWE-agent 5.71% $6.17
DeepSeek-V4-Pro max Codex 3.81% $5.33
GPT-5.6 Luna max Codex 3.33% $1.82

3 models have held number one since its release in Aug 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

Anthropic OpenAI Google DeepMind
Own leaderboard 3 leaders
GPT-6 Astra 68.1%
Lab claims 3 leaders
Gemini 4 Carbon 64.8%

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 Reported by the lab in its model card 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 GPT-5.6 Sol Claude Opus 4.8 DeepSeek-V4-Pro 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) GPT-6 Astra Claude Fable 5.1

4 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Anthropic
OpenAI
Google DeepMind
Moonshot AI
DateLaunch postScoreAs the post writes it
28 Oct 2026 Gemini 4 Carbon 64.8 Terminal-Bench-Science 0.1 (AA)
26 Oct 2026 GPT-6.1 Terra Terminal-Bench-Science 0.1, max reasoning (chart)
24 Oct 2026 Kimi K3.1 Pro 38.1 Terminal-Bench-Science 0.1 (Claude Code)
13 Oct 2026 Gemini 4 Argon 57.6 Terminal-Bench Science 0.1
29 Sep 2026 GPT-6.1 Sol 57 Terminal-Bench Science 0.1 (GPT-6.1 Sol Max)
22 Sep 2026 Claude Opus 5.5 58.7 Agentic scientific research Terminal-Bench-Science 0.1
3 Sep 2026 GPT-6 Astra 64.6 Terminal-Bench Science 0.1
1 Sep 2026 Claude Fable 5.1 52.6 Agentic scientific research Terminal-Bench-Science 0.1

5 system cards report it, from 2 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
13 Oct 2026 Gemini 4 Argon model card 57.6 Terminal-Bench Science 0.1 (results table, p.5) Capabilities
7 Oct 2026 Claude Haiku 5.5 system card 20.6 Terminal-Bench-Science 0.1, max effort, safeguards on with no fallback (8.5) Capabilities
28 Sep 2026 Claude Sonnet 5.5 system card 59.9 Terminal-Bench-Science 0.1, max effort, Claude Code --bare mode (8.6) Capabilities
22 Sep 2026 Claude Opus 5.5 system card 58.7 Terminal-Bench-Science 0.1, max effort, Claude Code --bare mode, safeguards with fallback, 10 trials per task (Table 8.1.A, 8.6) Capabilities
1 Sep 2026 Claude Fable 5.1 system card 52.6 Terminal-Bench-Science 0.1, Claude Code --bare mode, max effort, 10 trials per task (Table 8.1.A, 8.7) Capabilities

More in Science and reasoning

All 88 in Science and reasoning
  • GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics. Gemini 4 Argon 96.4%
  • 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 Opus 5 90.1%
  • 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%
  • SGI-Bench In use Models work through more than 1,000 expert samples in four scientist tasks: deep research, idea generation, dry and wet experiments, and experimental reasoning. Intern-S2-397B 51.9%