Terminal-Bench-Science leaderboard
As of November 8, 2026, GPT-6 Astra (OpenAI) leads Terminal-Bench-Science at 68.1%, ahead of Claude Opus 5.5 at 58.7% and Gemini 4 Argon at 57.6%. 9 systems ranked. Every score here is published by the model's own lab; each row links to the page it comes from.
What is Terminal-Bench-Science? Terminal-Bench-Science tests AI agents on real scientific research workflows, contributed by domain experts, carried out in a terminal: data analysis, simulation, theorem proving, image reconstruction, model fitting. Made by Led by Stanford University researchers with the Terminal-Bench team and domain experts, August 2026.
Top models on Terminal-Bench-Science, November 2026
| # | Model | Score | $ / task | Price / 1M | Context | Source |
|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | max effort | $23.80 | $10 / $50 | 1.05M | Vendor-reported OpenAI |
| 2 | Claude Opus 5.5 Anthropic | adaptive thinking, max effort | — | $4 / $20 | 1M | Vendor-reported Anthropic |
| 3 | Gemini 4 Argon Google DeepMind | as reported | — | $2 / $10 | 1M | Vendor-reported Google DeepMind |
| 4 | GPT-6.1 Sol OpenAI | max effort | $5.47 | $2 / $10 | 1.05M | Vendor-reported OpenAI |
| 5 | Claude Fable 5.1 Anthropic | as reported | — | $10 / $50 | 1M | Vendor-reported Anthropic |
| 6 | Kimi K3.1 Pro Moonshot (Kimi) | max effort | — | $2 / $10 | 1M | Vendor-reported Moonshot AI |
| 7 | Claude Opus 5 Anthropic | as reported | — | $5 / $25 | 1M | Vendor-reported Anthropic |
| 8 | GPT-6 Sol OpenAI | max effort | $12.18 | $2 / $10 | 1.05M | Vendor-reported OpenAI |
| 9 | GPT-5.6 Sol OpenAI | as reported | — | $4 / $20 | 1.05M | Vendor-reported OpenAI |
Terminal-Bench-Science: score against price
What more reasoning buys
Some publishers ran the same model at several reasoning-effort levels and printed the score and the cost per task for each. Higher effort spends more tokens per task; this is what it bought on Terminal-Bench-Science.
| Model | Score by effort level ($ per task) | Source |
|---|---|---|
| GPT-6 Astra | low 55.4% $11.41 medium 57.4% $12.34 high 62% $14.95 xhigh 60.9% $15.76 max 68.1% $23.80 | OpenAI |
| GPT-6.1 Sol | low 43.7% $1.79 medium 47.6% $2.34 high 51.1% $2.76 xhigh 53.7% $2.89 max 57% $5.47 | OpenAI |
| GPT-6 Sol | low 9.2% $3.00 medium 14.5% $4.41 high 14.6% $4.63 xhigh 25.3% $6.77 max 27.6% $12.18 | OpenAI |
Can you trust Terminal-Bench-Science scores?
No independent leaderboard publishes Terminal-Bench-Science results for these models yet, so every number on this page is the lab's own report of its own model. Labs choose the harness, the effort level and the subset they report, and they rarely publish failures. Read the setting under each score before comparing two rows.
- Tasks were calibrated against the newest frontier models, so absolute rates run well below Terminal-Bench 3.0.
- Only 70 of 920 proposals made it into 0.1, so coverage of each scientific field is thin.
- It is a versioned, continuous benchmark; scores belong to version 0.1.
How Terminal-Bench-Science works
The task
Each task is a research workflow an expert contributed; the agent runs it in a terminal environment and delivers an artifact that can be checked.
The score
Task-specific tests grade the artifacts; the headline is the resolution rate.
Test set
70 tasks (version 0.1), as published by the authors.
Made by
Led by Stanford University researchers with the Terminal-Bench team and domain experts, first published August 2026.
Official sources
- www.terminal-bench-science.ai
- Definition read on 2026-11-08
Sources and freshness
9 published scores from 5 sources. Each was read on its source page; the oldest reading on this page is from November 8, 2026. Prices beside each score come from our model price index and move when a vendor reprices.
- Vendor-reported Anthropic
- Vendor-reported OpenAI
- Vendor-reported Google DeepMind
- Vendor-reported Moonshot AI
- Vendor-reported OpenAI
Terminal-Bench-Science FAQ
Answers derived from the 9 published scores on this page and the current price index.