Terminal-Bench Science team
Terminal-Bench Science 0.1 — Life Sciences
AI agents tackle 19 life-science research workflows, the official Life Sciences subset of the 70-task benchmark.
biologymedicineneuroscienceecologycodetexttabularimage
Status
curated
Saturation
unknown
Reproduction
not yet
Published results
17 results
Published results
Official Life Sciences subset: 19 of 70 tasks, with three trials per task (57 trials per configuration). These are the authors’ published evaluations, not BioInspect runs. ± values are source-reported standard errors, not 95% confidence intervals. Download results JSON
| Model | Score | Variant | Source |
|---|---|---|---|
| GPT-6 Astra (max) · Codex | 59.6% resolution rate ± 6.5 pp SE 34/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-22 |
| Opus 5.5 (max) · Claude Code | 54.4% resolution rate ± 6.6 pp SE 31/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-23 |
| Fable 5.1 (max) · Claude Code | 38.6% resolution rate ± 6.4 pp SE 22/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Opus 5 (max) · Claude Code | 29.8% resolution rate ± 6.1 pp SE 17/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| GPT-5.6 Sol (max) · Codex | 17.5% resolution rate ± 5.0 pp SE 10/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Fable 5 (max) · Claude Code | 17.5% resolution rate ± 5.0 pp SE 10/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Gemini 3.8 Flash (high) · mini-SWE-agent | 14.0% resolution rate ± 4.6 pp SE 8/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Grok 4.7 (xhigh) · Grok Build | 12.3% resolution rate ± 4.3 pp SE 7/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-21 |
| DeepSeek V4.1 Flash (max) · Codex | 10.5% resolution rate ± 4.1 pp SE 6/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| GLM 5.3 (max) · Claude Code | 10.5% resolution rate ± 4.1 pp SE 6/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Gemini 3.7 Flash (high) · mini-SWE-agent | 8.8% resolution rate ± 3.7 pp SE 5/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Opus 4.8 (max) · Claude Code | 7.0% resolution rate ± 3.4 pp SE 4/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| GPT-5.6 Terra (max) · Codex | 7.0% resolution rate ± 3.4 pp SE 4/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Kimi K3 (max) · Claude Code | 7.0% resolution rate ± 3.4 pp SE 4/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| Grok 4.6 (xhigh) · Grok Build | 7.0% resolution rate ± 3.4 pp SE 4/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| GPT-5.6 Luna (max) · Codex | 1.8% resolution rate ± 1.7 pp SE 1/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
| DeepSeek V4 Pro (max) · Codex | 0.0% resolution rate ± 0.0 pp SE 0/57 trials | v0.1 · Life Sciences · 19 tasks × 3 trials | Terminal-Bench Science official Life Sciences results 2026-09-16 |
Sources
- external Terminal-Bench Science official Life Sciences results
- official announcement Benchmark design and three independent trials per task
- official repository Task definitions; execution uses Harbor, not Inspect
- official source JSON Domain aggregates and trial outcomes