Why this benchmark is useful
For source-heavy students and researchers this is the live science-agent discriminator: real lab-style workflows, still far from ceiling (best published 0.1 row is 30%).
Terminal-Bench-Science
TB-Science
Whether an agent can finish real scientific research workflows in a terminal: 70 expert-curated tasks across life, physical, Earth, mathematical, and engineering sciences, graded on concrete artifacts (analyses, simulations, proofs, code, data products).
What this does not measure
- The model alone — every published 0.1 row names an agent harness (Claude Code, Codex, Grok Build). Harness choice moves the score.
- Hypothesis generation or paper writing — tasks are executable workflows with hidden checkers, not open-ended discovery.
- Terminal-Bench 2.1 software-engineering skill — TB-Science is a separate, harder science suite. Do not mix 2.1 coding pass rates with 0.1 science resolution.
Coverage map
- Task family
- Agentic
- Format
- 70 containerized Harbor tasks. Agents get a terminal, files, and tools; graders check the produced artifact with task-specific tests. Three independent trials per task in the 0.1 announcement table.
- Scoring
- Resolution rate across 70 tasks (pass if the hidden checker accepts the artifact). VerdictPal rows use the 27 Aug 2026 announcement table, one row per model at its published harness.
- Maintainer
- Stanford / Laude Institute / Terminal-Bench
How to read the scores
Read resolution rate with the harness named on the row. Claude Opus 5 + Claude Code leads the 0.1 table at 30.0%; GPT-5.6 Sol + Codex is 22.4%. GLM-5.3 is the strongest open-weight row at 8.1%. Do not compare these to Terminal-Bench 2.1 coding headlines.
What it does not cover
- The model alone — every published 0.1 row names an agent harness (Claude Code, Codex, Grok Build). Harness choice moves the score.
- Hypothesis generation or paper writing — tasks are executable workflows with hidden checkers, not open-ended discovery.
- Terminal-Bench 2.1 software-engineering skill — TB-Science is a separate, harder science suite. Do not mix 2.1 coding pass rates with 0.1 science resolution.
Scores
9 rowsEvidence ledger
99 rows
30%best score
9source-checked
1sources
2026-08-27source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- Claude Opus 530% · Terminal-Bench-Science 0.1 announcement
- GPT-5.6 Sol22.4% · Terminal-Bench-Science 0.1 announcement
- Claude Fable 521.4% · Terminal-Bench-Science 0.1 announcement
- Claude Opus 4.810.5% · Terminal-Bench-Science 0.1 announcement
- GPT-5.6 Terra8.6% · Terminal-Bench-Science 0.1 announcement
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | Claude Opus 5AnthropicOfficial 0.1 table: Claude Code harness, three trials × 70 tasks. Do not mix with Terminal-Bench 2.1. | 2026-07-24 | 30% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 2 | GPT-5.6 SolOpenAIOfficial 0.1 table: Codex harness. Matches Fable 5 resolution at less than a third of the announced eval cost. | 2026-07-09 | 22.4% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 3 | Claude Fable 5Anthropic | 2026-06-09 | 21.4% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 4 | Claude Opus 4.8Anthropic | 2026-05-28 | 10.5% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 5 | GPT-5.6 TerraOpenAI | 2026-07-09 | 8.6% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 6 | GLM-5.3ZhipuStrongest open-weight row on the 0.1 announcement table. | 2026-08-14 | 8.1% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 7 | Kimi K3Moonshot | 2026-07-16 | 7.1% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 7 | Grok 4.6xAI | 2026-08-12 | 7.1% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
| 9 | GPT-5.6 LunaOpenAI | 2026-07-09 | 3.3% | Terminal-Bench-Science 0.1 announcement2026-08-27 | Source-checked |
What it covers
Scores from the Terminal-Bench-Science 0.1 announcement (27 Aug 2026): three trials per task, 70 tasks. Work on 0.2 is open with a 5 Oct 2026 PR deadline. Official Terminal-Bench 4.0 is a different, software-engineering suite — not this page.
Where it breaks down
Human expert time is research-workflow scale. Published 0.1 resolution rates cluster at or below 30%.
Receipts
Sources and further reading
- Terminal-Bench-Science 0.1 announcement· checked 2026-09-01
- Live leaderboard· checked 2026-09-01
- GitHub task suite· checked 2026-09-01
- Announcement
- Leaderboard
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does Terminal-Bench-Science measure?
Whether an agent can finish real scientific research workflows in a terminal: 70 expert-curated tasks across life, physical, Earth, mathematical, and engineering sciences, graded on concrete artifacts (analyses, simulations, proofs, code, data products).
What does a high Terminal-Bench-Science score not prove?
A strong Terminal-Bench-Science result says nothing about:
- The model alone — every published 0.1 row names an agent harness (Claude Code, Codex, Grok Build). Harness choice moves the score.
- Hypothesis generation or paper writing — tasks are executable workflows with hidden checkers, not open-ended discovery.
- Terminal-Bench 2.1 software-engineering skill — TB-Science is a separate, harder science suite. Do not mix 2.1 coding pass rates with 0.1 science resolution.
How is Terminal-Bench-Science scored?
Resolution rate across 70 tasks (pass if the hidden checker accepts the artifact). VerdictPal rows use the 27 Aug 2026 announcement table, one row per model at its published harness. Task format: 70 containerized Harbor tasks. Agents get a terminal, files, and tools; graders check the produced artifact with task-specific tests. Three independent trials per task in the 0.1 announcement table.
Is Terminal-Bench-Science saturated?
Terminal-Bench-Science is currently marked Active in the atlas. Ceiling context: Human expert time is research-workflow scale. Published 0.1 resolution rates cluster at or below 30%.
Can Terminal-Bench-Science results be contaminated by training data?
Contamination risk for Terminal-Bench-Science is graded Low contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.