BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases7

Terminal-Bench Science leaderboard

As of 6 Nov 2026, the top score on Terminal-Bench Science is 62.9% by GPT-6 Astra (OpenAI), ahead of GPT-6.1 Sol at 52.9%, across 44 model configurations with a published result (Vals AI, result dated 5 Nov 2026). Scientists wrote multi-step research workflows, from data processing to analysis, for an agent to complete in a shell. Terminal-Bench Science is shown for reference only and does not count toward the BenchLeader Index.

Key facts (as of 6 Nov 2026)

  • GPT-6 Astra holds the top Terminal-Bench Science score at 62.9% (Vals AI, 5 Nov 2026).
  • GPT-6.1 Sol is second on Terminal-Bench Science at 52.9% (Vals AI, 5 Nov 2026).
  • Terminal-Bench Science: Expert-authored scientific research workflows in a terminal. Run by Vals AI (Vals AI).
  • 44 model configurations have a published Terminal-Bench Science result on BenchLeader (BenchLeader, 6 Nov 2026).
  • Terminal-Bench Science does not feed the BenchLeader Index; it is shown for reference (BenchLeader methodology).

Expert-authored scientific research workflows in a terminal. Run by Vals AI.

Published by
Vals AI
Category
Agents & tools
Index weight
Reference only
Models
44
Data as of
6 Nov 2026

Vals AI (vals.ai).

What the test looks like

Scientists wrote multi-step research workflows, from data processing to analysis, for an agent to complete in a shell.

How it is scored

Percent of workflows completed, run by Vals AI.

What to keep in mind

Domain knowledge and tooling both matter; few models measured.

Top 15
  1. 1GPT-6 Astra (max)62.9%
  2. 2GPT-6.1 Sol (max)52.9%
  3. 3Claude Opus 5.547.1%
  4. 4Claude Sonnet 5.547.1%
  5. 5Gemini 4 Argon (high)44.3%
  6. 6Claude Fable 5.140.0%
  7. 7Muse Wonder Full (max)37.1%
  8. 8Kimi K3.1 Pro (max)31.4%
  9. 9GPT-6 Sol (max)30.0%
  10. 10Claude Opus 527.1%
  11. 11Muse Wonder (max)22.9%
  12. 12GPT-5.6 Sol (max)20.0%
  13. 13Claude Fable 515.7%
  14. 14Qwen3.8 Max (max)12.9%
  15. 15Grok 4.6 (high)11.4%
44 of 44
#
1GPT-6 Astramax62.9%67.8$20.802026-11-05
2GPT-6.1 Solmax52.9%67.3$3.442026-11-05
3Claude Opus 5.547.1%63.4$19.122026-11-05
4Claude Sonnet 5.547.1%–$19.572026-11-05
5Gemini 4 Argonhigh44.3%66.0$44.902026-11-05
6Claude Fable 5.140.0%66.4$38.012026-11-05
7Muse Wonder Fullmax37.1%62.9$22.812026-11-05
8Kimi K3.1 Promax31.4%–$8.582026-11-05
9GPT-6 Solmax30.0%63.3$5.822026-11-05
10Claude Opus 527.1%65.2$32.542026-11-05
11Muse Wondermax22.9%62.3$19.612026-11-05
12GPT-5.6 Solmax20.0%65.4$7.322026-11-05
13Claude Fable 515.7%66.5$51.992026-11-05
14Qwen3.8 Maxmax12.9%58.8$16.682026-11-05
15Grok 4.6high11.4%60.5$4.822026-11-05
16Muse Spark 1.3max10.0%62.0$6.092026-11-05
17GPT-5.6 Terramax10.0%60.7$5.192026-11-05
18Grok 4.7xhigh10.0%58.3$17.042026-11-05
19Claude Haiku 5.510.0%–$4.522026-11-05
20Gemini 3.8 Flashhigh8.6%62.0$5.642026-11-05
21Gemini 3.7 Flashhigh7.1%62.2$7.192026-11-05
22Gemini 3.5 Flashhigh5.7%62.5$3.802026-11-05
23GLM 5.3max5.7%59.0$15.632026-11-05
24Claude Sonnet 55.7%55.9$30.302026-11-05
25Ember-15.7%–$30.762026-11-05
26Gemini 3.6 Flashhigh4.3%61.2$4.392026-11-05
27Claude Opus 4.84.3%60.0$23.142026-11-05
28Muse Spark 1.3xhigh4.3%58.5$4.412026-11-05
29DeepSeek V4 Flashhigh4.3%57.3$1.102026-11-05
30DeepSeek V4 Promax4.3%57.0$4.632026-11-05
31GPT-6 Lunamax4.3%56.7$0.2412026-11-05
32MiMo-V2.6-Flash4.3%55.0$0.2442026-11-05
33NewInkling 2 Small2.9%57.4$5.552026-11-05
34MiMo-V2.6-Pro2.9%57.3$0.6212026-11-05
35Step 5 Preview2.9%54.9$4.252026-11-05
36GLM 5.3 Flashmax2.9%54.2$0.5242026-11-05
37Kimi K3max1.4%62.1$21.302026-11-05
38Gemini 3.1 Prohigh1.4%60.1$5.912026-11-05
39Hy41.4%55.7$2.142026-11-05
40Beam 1.01.4%45.8–2026-11-05
41DeepSeek V4.1 Flashmax0.0%59.0$0.4732026-11-05
42GPT-5.6 Lunamax0.0%56.3$0.5812026-11-05
43Qwen3.8 27Bxhigh0.0%52.5$11.142026-11-05
44Mercury 2.5high0.0%–$0.1782026-11-05

More agents & tools benchmarks · All LLM benchmarks

Related benchmarks

Top head-to-heads on Terminal-Bench Science

Cite as: BenchLeader, “Terminal-Bench Science leaderboard”, data as of 6 Nov 2026, https://www.benchleader.com/benchmarks/vals_terminal_bench_science.

Terminal-Bench Science: questions

What does Terminal-Bench Science measure?
Scientists wrote multi-step research workflows, from data processing to analysis, for an agent to complete in a shell. Scores are reported in percent of tasks solved; higher is better.
What is the top score on Terminal-Bench Science?
As of 6 Nov 2026, the highest score on Terminal-Bench Science is 62.9% by GPT-6 Astra (result dated 5 Nov 2026), ahead of GPT-6.1 Sol at 52.9%. Scores come from Vals AI and are re-read daily.
Which AI model leads Terminal-Bench Science?
GPT-6 Astra leads Terminal-Bench Science with 62.9% as of 6 Nov 2026 (result dated 5 Nov 2026), ahead of GPT-6.1 Sol at 52.9%.
How many models have Terminal-Bench Science results?
44 model configurations have a Terminal-Bench Science result on BenchLeader, all taken from Vals AI.
Who runs Terminal-Bench Science and how often is it updated?
Terminal-Bench Science is published by Vals AI. BenchLeader re-reads the published results once a day and records the date each result was published.
Does Terminal-Bench Science count toward the BenchLeader Index?
No. Terminal-Bench Science is shown for reference but left out of the composite index.