Terminal-Bench-Science (Vals AI) leaderboard
- GPT-6 Astra leads
- 62.9%
- models measured
- 39
- lead over #2
- +10.0 pts
- median
- 5.7%
Leader on Terminal-Bench-Science (Vals AI) · read October 22
On October 22, 2026, 39 of our catalog's models had a Terminal-Bench-Science (Vals AI) result, and OpenAI GPT-6 Astra led with 62.9%, ahead of OpenAI GPT-6.1 Sol at 52.9%.
Scientists' research workflows, done by an agent in a terminal. Share completed.
Role in the index
Terminal-Bench-Science (Vals AI) is shown as published and does not feed any of our indexes. Read the methodology
| # | Model | Result | Cost per test | Duration, as printed | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 AstraOpenAI | 62.9% | $20.80 | 1h 49m | 85.1 | $50.00 | |
| 2 | GPT-6.1 SolOpenAI | 52.9% | $3.44 | 1h 23m | 82.4 | $10.00 | |
| 3 | Claude Opus 5.5Anthropic | 47.1% | $19.12 | 2h 28m | 88.8 | $20.00 | |
| 4 | Claude Sonnet 5.5Anthropic | 47.1% | $32.52 | 2h 46m | 87.3 | $10.00 | |
| 5 | Gemini 4 ArgonGoogle | 44.3% | $44.90 | 1h 46m | 81.0 | $10.00 | |
| 6 | Claude Fable 5.1Anthropic | 40.0% | $38.01 | 2h 36m | 84.2 | $50.00 | |
| 7 | GPT-6 SolOpenAI | 30.0% | $5.82 | 1h 22m | 76.0 | $10.00 | |
| 8 | Claude Opus 5Anthropic | 27.1% | $32.54 | 2h 12m | 83.0 | $25.00 | |
| 9 | Muse WonderMeta | 22.9% | $19.61 | 1h 33m | 71.9 | $20.00 | |
| 10 | GPT-5.6 SolOpenAI | 20.0% | $7.32 | 1h 44m | 73.5 | $20.00 | |
| 11 | Claude Fable 5Anthropic | 15.7% | $51.99 | 1h 59m | 79.3 | $50.00 | |
| 12 | Qwen3.8 Max (0902)Qwen | 12.9% | $16.68 | 3h 24m | 62.4 | $6.00 | |
| 13 | Grok 4.6xAI | 11.4% | $4.82 | 1h 18m | 65.7 | $6.00 | |
| 14 | Claude Haiku 5.5Anthropic | 10.0% | $4.52 | 1h 34m | 70.7 | $0.50 | |
| 15 | GPT-5.6 TerraOpenAI | 10.0% | $5.19 | 2h 04m | 63.9 | $12.00 | |
| 16 | Grok 4.7xAI | 10.0% | $17.04 | 1h 33m | 66.7 | $6.00 | |
| 17 | Gemini 3.8 FlashGoogle | 8.6% | $5.64 | 55m 11s | 61.6 | $3.75 | |
| 18 | Gemini 3.7 FlashGoogle | 7.1% | $7.19 | 1h 20m | 58.0 | $3.75 | |
| 19 | Claude Sonnet 5Anthropic | 5.7% | $30.30 | 3h 06m | 57.7 | $10.00 | |
| 20 | Ember-1Fireworks | 5.7% | $30.76 | 3h 23m | 52.6 | $15.00 | |
| 21 | Gemini 3.5 FlashGoogle | 5.7% | $3.80 | 1h 01m | 39.8 | $9.00 | |
| 22 | GLM 5.3Z.ai | 5.7% | $15.63 | 2h 45m | 70.6 | $4.40 | |
| 23 | Claude Opus 4.8Anthropic | 4.3% | $23.14 | 1h 55m | 65.0 | $25.00 | |
| 24 | DeepSeek V4 Flash 0731DeepSeek | 4.3% | $1.10 | 1h 34m | 51.8 | $0.396 | |
| 25 | DeepSeek V4 Pro 0813DeepSeek | 4.3% | $4.63 | 2h 16m | 56.1 | $1.98 | |
| 26 | Gemini 3.6 FlashGoogle | 4.3% | $4.39 | 1h 52m | 62.5 | $3.75 | |
| 27 | Muse Spark 1.3Meta | 4.3% | $4.41 | 1h 08m | 57.4 | $4.25 | |
| 28 | GPT-6 LunaOpenAI | 4.3% | $0.24 | 1h 16m | 57.3 | $0.50 | |
| 29 | MiMo-V2.6-FlashXiaomi | 4.3% | $0.24 | 3h 59m | 54.7 | $0.28 | |
| 30 | Step 5 PreviewStepFun | 2.9% | $4.25 | 2h 30m | 61.1 | $2.70 | |
| 31 | MiMo-V2.6-ProXiaomi | 2.9% | $0.62 | 3h 12m | 65.4 | $0.87 | |
| 32 | GLM 5.3 FlashZ.ai | 2.9% | $0.52 | 3h 28m | 57.7 | $0.50 | |
| 33 | Gemini 3.1 Pro PreviewGoogle | 1.4% | $5.91 | 43m 58s | 32.3 | $12.00 | |
| 34 | Kimi K3Moonshot AI | 1.4% | $21.30 | 4h 49m | 62.3 | $15.00 | |
| 35 | Hy4 previewTencent | 1.4% | $2.14 | 3h 48m | 44.3 | $2.501 | |
| 36 | DeepSeek V4.1 FlashDeepSeek | 0.0% | $0.47 | 2h 57m | 62.5 | $0.84 | |
| 37 | Mercury 2.5Inception | 0.0% | $0.18 | 28m 14s | 1.6 | $0.15 | |
| 38 | GPT-5.6 LunaOpenAI | 0.0% | $0.58 | 2h 03m | 58.0 | $1.20 | |
| 39 | Qwen3.8 27BQwen | 0.0% | $11.14 | 4h 02m | 74.2 | $2.55 |
Only models in our catalog are listed, and each appears once with its best published run. 137 rows were read from the source on October 22, 2026, and rows that match no catalog model are left out. Cost per test is what Vals AI reports one test cost on the run behind the result, in US dollars. Every other column is a figure the board publishes for the same run.