TerraBench - NumScore: leaderboard
NumScore on TerraBench (2026), an executable Earth-science benchmark for tool-using agents: tolerance-aware partial credit for numeric answers that halves for each extra tolerance width of error, with missing answers scoring 0, on a 0 to 100 scale. Higher is better.
Metric: NumScore (%): tolerance-aware partial credit for numeric answers, halving for each extra tolerance width of error (missing or unparseable answers score 0), on the 403 executable TerraBench tasks (Fundamentals, Simulator-Grounded and Document-Grounded Verification tracks across eight Earth-science domains), each model driving the benchmark's TerraAgent ReAct harness with its 77 scientific tools; higher is better. Source: arxiv.org. Saturation forecast: Around July 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 28.44 | #161 |
| 2 | GPT-5.5 | 26.45 | #49 |
| 3 | Claude Haiku 4.5 | 21.19 | #568 |
| 4 | Gemini 3.1 Pro (Preview) | 17.07 | #94 |
| 5 | GPT-5.4 | 12.29 | #123 |
| 6 | Gemini 2.5 Flash | 7.7 | #453 |
| 7 | Qwen 3.5 35B A3B | 7.49 | #482 |
| 8 | Gemma 4 26B A4B | 4.45 | #615 |
| 9 | Qwen 3 14B | 4.3 | #1233 |
| 10 | Gemma 4 E4B | 3.04 | #2214 |
| 11 | Mistral 7B Instruct (v0.3) | 2.41 | #6653 |
| 12 | Qwen 3 8B | 2.06 | #1830 |
| 13 | InternVL3-8B | 1.7 | #1688 |
| 14 | Qwen 3.5 9B | 1.3 | #803 |
| 15 | Qwen 3 1.7B | 1.15 | #5520 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=terrabench-numscore · How It Works · Data refreshed daily, snapshot 2026-11-13.