Terminal-Bench Science leaderboard

Expert-written scientific research workflows, run by an agent in a terminal. GPT-6 Astra leads with 62.86%.

Models
42
Leader
GPT-6 Astra
Source
Vals AI
As of
Oct 24, 2026
  1. 1GPT-6 Astra62.9%
  2. 2GPT-6.1 Sol52.9%
  3. 3Claude Opus 5.547.1%
  4. 4Claude Sonnet 5.547.1%
  5. 5Gemini 4 Argon44.3%
  6. 6Claude Fable 5.140.0%
  7. 7Muse Wonder Full37.1%
  8. 8GPT-6 Sol30.0%
  9. 9Claude Opus 527.1%
  10. 10Muse Wonder22.9%
Terminal-Bench Science, accuracy against cost per test
Open weightsProprietaryBest for the price
0%18%35%53%70%$.1$1$10$100Cost per test, log scale →GPT-6 AstraGPT-6.1 SolGPT-6 LunaMercury 2.5

Full leaderboard

RankModelLabAccuracyCost / testHarnessPrice in / out
1GPT-6 AstraOpenAI62.86%$20.80Mini-SWE-agent$10 / $50
2GPT-6.1 SolOpenAI52.86%$3.44Mini-SWE-agent$2 / $10
3Claude Opus 5.5Anthropic47.14%$19.12Mini-SWE-agent$4 / $20
4Claude Sonnet 5.5Anthropic47.14%$32.52Mini-SWE-agent$2 / $10
5Gemini 4 ArgonGoogle44.29%$44.90Mini-SWE-agent$2 / $10
6Claude Fable 5.1Anthropic40.00%$38.01Mini-SWE-agent$10 / $50
7Muse Wonder FullMeta37.14%$22.81Mini-SWE-agent—
8GPT-6 SolOpenAI30.00%$5.82Mini-SWE-agent$2 / $10
9Claude Opus 5Anthropic27.14%$32.54Mini-SWE-agent$5 / $25
10Muse WonderMeta22.86%$19.61Mini-SWE-agent$4 / $20
11GPT-5.6 SolOpenAI20.00%$7.32Mini-SWE-agent$4 / $20
12Claude Fable 5Anthropic15.71%$51.99Mini-SWE-agent$10 / $50
13Qwen3.8-MaxAlibaba12.86%$16.68Mini-SWE-agent$2 / $6
14Grok 4.6xAI11.43%$4.82Mini-SWE-agent$2 / $6
15Muse Spark 1.3 MaxMeta10.00%$6.09Mini-SWE-agent—
16Grok 4.7xAI10.00%$17.04Mini-SWE-agent$2 / $6
17Claude Haiku 5.5Anthropic10.00%$4.52Mini-SWE-agent$0.10 / $0.50
18GPT-5.6 TerraOpenAI10.00%$5.19Mini-SWE-agent$2 / $12
19Gemini 3.8 FlashGoogle8.57%$5.64Mini-SWE-agent$0.82 / $4.13
20Gemini 3.7 FlashGoogle7.14%$7.19Mini-SWE-agent$0.75 / $3.75
21Gemini 3.5 FlashGoogle5.71%$3.80Mini-SWE-agent$1.50 / $9
22GLM-5.3Z.ai5.71%$15.63Mini-SWE-agent$0.079 / $1.40
23Claude Sonnet 5Anthropic5.71%$30.30Mini-SWE-agent$2 / $10
24Ember-1Fireworks AI5.71%$30.76Mini-SWE-agent$3 / $15
25Muse Spark 1.3Meta4.29%$4.41Mini-SWE-agent$1.25 / $4.25
26GPT-6 LunaOpenAI4.29%$0.24Mini-SWE-agent$0.10 / $0.50
27DeepSeek-V4-FlashDeepSeek4.29%$1.10Mini-SWE-agent$0.042 / $0.18
28Gemini 3.6 FlashGoogle4.29%$4.39Mini-SWE-agent$0.75 / $3.75
29Claude Opus 4.8Anthropic4.29%$23.14Mini-SWE-agent$5 / $25
30DeepSeek-V4-ProDeepSeek4.29%$4.63Mini-SWE-agent$0.66 / $1.98
31MiMo-V2.6-FlashXiaomi4.29%$0.24Mini-SWE-agent$0.14 / $0.28
32Step 5 PreviewStepFun2.86%$4.25Mini-SWE-agent$1 / $2.70
33MiMo-V2.6-ProXiaomi2.86%$0.62Mini-SWE-agent$0.43 / $0.87
34GLM-5.3-FlashZ.ai2.86%$0.52Mini-SWE-agent$0.15 / $0.50
35Gemini 3.1 Pro PreviewGoogle1.43%$5.91Mini-SWE-agent$2 / $12
36Beam 1.0Reflection AI1.43%N/AMini-SWE-agent—
37Hy4 previewTencent1.43%$2.14Terminus 2$0.75 / $2.25
38Kimi K3Moonshot AI1.43%$21.30Mini-SWE-agent$0.24 / $11
39Mercury 2.5Inception0.00%$0.18Mini-SWE-agent$0.04 / $0.15
40GPT-5.6 LunaOpenAI0.00%$0.58Mini-SWE-agent$0.20 / $1.20
41DeepSeek-V4.1-Flash-20260910DeepSeek0.00%$0.47Mini-SWE-agent$0.15 / $0.60
42Qwen3.8-27BAlibaba0.00%$11.14Mini-SWE-agent$0.20 / $2.50

Related leaderboards

Sources

  • Vals AI, as of Oct 24, 2026
  • Provider list prices, as of Oct 24, 2026

Every number on this page comes from these sources; models without a published number show none. Download the data. Cite as “Nexus Explorer, nexus.inc/models”.