Terminal-Bench-Science: leaderboard
Terminal-Bench-Science evaluates terminal agents on long-horizon scientific research and engineering tasks in isolated environments. The evaluated model operates from text inputs. The pinned metric is resolution rate. Source: BenchmarkList’s attributed result table.
Metric: Resolution rate (%). Source: benchmarklist.com. Saturation forecast: Around July 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 68.1 | #6 (GPT-6 Astra) |
| 2 | Claude Opus 5.5 (Max) | 63.3 | #3 (Claude Opus 5.5) |
| 3 | Claude Fable 5.1 (Max) | 40 | #10 (Claude Fable 5.1) |
| 4 | Kimi K3.1 Pro (Max) | 38.1 | #17 (Kimi K3.1 Pro) |
| 5 | DeepSeek V4.1 Flash (Max) | 15.7 | #37 (DeepSeek V4.1 Flash) |
| 6 | GLM-5.3 (Max) | 8.1 | #44 (GLM-5.3) |
No result here: #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=terminal-bench-science · How It Works · Data refreshed daily, snapshot 2026-11-05.