Share promptsSign inCreate team library
← All benchmarks

Agents & tool use AI benchmarks

Multi-step tasks an agent completes alone in a terminal. 2 public leaderboards below.

Terminal-Bench 4.0

Hard, multi-step tasks an agent must complete on its own in a command-line environment.

Run by Vals AI (benchmark by Terminal-Bench team). Leaderboard as of 19 Oct 2026 · retrieved 21 Oct 2026. Showing top 25 of 52 models listed.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • Meta
  • DeepSeek
  • Alibaba
  • Z.ai
  • MiniMax
  • Mistral
  • Xiaomi
  • Other
Terminal-Bench 4.0: Score vs Cost per test (USD)Scatter plot of 25 models. Leader on Score: Claude Opus 5.5 at 65.15%. Full values in the table below.$.1$1$10$10010%20%30%40%50%60%70%Cost per test (USD) (log scale) →↑ ScoreClaude Opus 4.8 (Anthropic): Cost per test (USD) $17, Score 23.23%AMiMo V2.6 Flash (Xiaomi): Cost per test (USD) $0.22, Score 24.24%XiMuse Spark 1.3 Max (Meta): Cost per test (USD) $6.65, Score 24.75%MMiniMax-M3.1-Flash (MiniMax): Cost per test (USD) $0.73, Score 25.25%MMMistral Large 4 (Mistral): Cost per test (USD) $10, Score 27.78%MiGrok 4.7 (xAI): Cost per test (USD) $18, Score 28.79%XMiMo V2.6 Pro (Xiaomi): Cost per test (USD) $0.50, Score 31.31%XiStep 5 Preview (Other): Cost per test (USD) $3.11, Score 31.82%·DeepSeek V4.1 Flash 1017 (DeepSeek): Cost per test (USD) $0.54, Score 31.82%DQwen 3.8 Max (Alibaba): Cost per test (USD) $11, Score 34.34%QClaude Haiku 5.5 (Anthropic): Cost per test (USD) $3.12, Score 36.36%AGLM 5.3 Flash (Z.ai): Cost per test (USD) $1.22, Score 37.37%ZGPT-5.6 Sol (OpenAI): Cost per test (USD) $7.98, Score 37.88%OGLM 5.3 (Z.ai): Cost per test (USD) $9.37, Score 38.89%ZClaude Fable 5 (Anthropic): Cost per test (USD) $30, Score 41.41%AGPT-6 Sol (OpenAI): Cost per test (USD) $5.79, Score 44.44%OWonder (Meta): Cost per test (USD) $15, Score 49.49%MClaude Opus 5 (Anthropic): Cost per test (USD) $19, Score 53.53%AGPT-6.1 Sol (OpenAI): Cost per test (USD) $1.72, Score 55.05%OGemini 4 Argon (Google): Cost per test (USD) $18, Score 57.58%GWonder Full (Meta): Cost per test (USD) $11, Score 58.08%MClaude Fable 5.1 (Anthropic): Cost per test (USD) $17, Score 58.08%AGPT-6 Astra (OpenAI): Cost per test (USD) $9.58, Score 59.60%OClaude Sonnet 5.5 (Anthropic): Cost per test (USD) $17, Score 64.14%AClaude Opus 5.5 (Anthropic): Cost per test (USD) $13, Score 65.15%AClaude Opus 5.5Claude Sonnet 5.5GPT-6 Astra
Terminal-Bench 4.0: top 15 by ScoreRanked bars. Leader: Claude Opus 5.5 at 65.15%. Full values in the table below.0%20%40%60%80%100%Claude Opus 5.5 (Anthropic): 65.15%AClaude Opus 5.565.15%Claude Sonnet 5.5 (Anthropic): 64.14%AClaude Sonnet 5.564.14%GPT-6 Astra (OpenAI): 59.60%OGPT-6 Astra59.60%Claude Fable 5.1 (Anthropic): 58.08%AClaude Fable 5.158.08%Wonder Full (Meta): 58.08%MWonder Full58.08%Gemini 4 Argon (Google): 57.58%GGemini 4 Argon57.58%GPT-6.1 Sol (OpenAI): 55.05%OGPT-6.1 Sol55.05%Claude Opus 5 (Anthropic): 53.53%AClaude Opus 553.53%Wonder (Meta): 49.49%MWonder49.49%GPT-6 Sol (OpenAI): 44.44%OGPT-6 Sol44.44%Claude Fable 5 (Anthropic): 41.41%AClaude Fable 541.41%GLM 5.3 (Z.ai): 38.89%ZGLM 5.338.89%GPT-5.6 Sol (OpenAI): 37.88%OGPT-5.6 Sol37.88%GLM 5.3 Flash (Z.ai): 37.37%ZGLM 5.3 Flash37.37%Claude Haiku 5.5 (Anthropic): 36.36%AClaude Haiku 5.536.36%
Score
Share of terminal tasks completed and verified by tests.
Cost per test (USD)
Average API cost per task, as published.
Terminal-Bench 4.0 scores
RankModelProviderScoreCost per test (USD)HarnessPrice in / out ($/1M tokens)
1 Claude Opus 5.5Anthropic65.15%$13Mini-SWE-agent$4 / $20
2 Claude Sonnet 5.5Anthropic64.14%$17Mini-SWE-agent$2 / $10
3 GPT-6 AstraOpenAI59.60%$9.58Mini-SWE-agent$10 / $50
4 Claude Fable 5.1Anthropic58.08%$17Mini-SWE-agent$10 / $50
5 Wonder FullMeta58.08%$11Mini-SWE-agent$4 / $20
6 Gemini 4 ArgonGoogle57.58%$18Mini-SWE-agent$4 / $20
7 GPT-6.1 SolOpenAI55.05%$1.72Mini-SWE-agent$2 / $10
8 Claude Opus 5Anthropic53.53%$19Mini-SWE-agent$5 / $25
9 WonderMeta49.49%$15Mini-SWE-agent$4 / $20
10 GPT-6 SolOpenAI44.44%$5.79Mini-SWE-agent$2 / $10
Show 15 more rows
RankModelProviderScoreCost per test (USD)HarnessPrice in / out ($/1M tokens)
11 Claude Fable 5Anthropic41.41%$30Mini-SWE-agent$10 / $50
12 GLM 5.3Z.ai38.89%$9.37Mini-SWE-agent$1.4 / $4.4
13 GPT-5.6 SolOpenAI37.88%$7.98Mini-SWE-agent$4 / $20
14 GLM 5.3 FlashZ.ai37.37%$1.22Mini-SWE-agent$0.075 / $0.25
15 Claude Haiku 5.5Anthropic36.36%$3.12Mini-SWE-agent$0.1 / $0.5
16 Qwen 3.8 MaxAlibaba34.34%$11Mini-SWE-agent$2 / $6
17 DeepSeek V4.1 Flash 1017DeepSeek31.82%$0.54Mini-SWE-agent$0.3 / $1.2
18 Step 5 PreviewOther31.82%$3.11Mini-SWE-agent$1 / $2.7
19 MiMo V2.6 ProXiaomi31.31%$0.50Mini-SWE-agent$0.435 / $0.87
20 Grok 4.7xAI28.79%$18Mini-SWE-agent$2 / $6
21 Mistral Large 4Mistral27.78%$10Mini-SWE-agent$1.36 / $4.18
22 MiniMax-M3.1-FlashMiniMax25.25%$0.73Mini-SWE-agent$0.15 / $0.6
23 Muse Spark 1.3 MaxMeta24.75%$6.65Mini-SWE-agent$1.25 / $4.25
24 MiMo V2.6 FlashXiaomi24.24%$0.22Mini-SWE-agent$0.14 / $0.28
25 Claude Opus 4.8Anthropic23.23%$17Mini-SWE-agent$5 / $25

Source: Vals AI leaderboard ↗ · Terminal-Bench project ↗

Top model here: Claude Opus 5.5 (Anthropic). Try prompts for Claude →

Terminal-Bench 2.1

The earlier, widely reported Terminal-Bench task set for command-line agents.

Run by Vals AI (benchmark by Terminal-Bench team). Leaderboard as of 28 Sep 2026 · retrieved 21 Oct 2026. Showing top 25 of 76 models listed.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • Meta
  • DeepSeek
  • Moonshot
  • Xiaomi
Terminal-Bench 2.1: Score vs Cost per test (USD)Scatter plot of 25 models. Leader on Score: Claude Opus 5.5 at 87.64%. Full values in the table below.$.01$.1$1$1070%75%80%85%90%Cost per test (USD) (log scale) →↑ ScoreClaude Opus 4.8 (Anthropic): Cost per test (USD) $2.41, Score 71.91%AMuse Spark 1.3 (Meta): Cost per test (USD) $0.66, Score 72.28%MGPT-6 Luna (OpenAI): Cost per test (USD) $0.030, Score 73.03%OGrok 4.7 (xAI): Cost per test (USD) $1.08, Score 73.41%XGemini 3.6 Flash (Google): Cost per test (USD) $1.17, Score 73.78%GGemini 3.5 Flash (Google): Cost per test (USD) $0.87, Score 74.16%GClaude Sonnet 5 (Anthropic): Cost per test (USD) $0.53, Score 74.53%ADeepSeek V4.1 Flash (DeepSeek): Cost per test (USD) $0.10, Score 74.53%DMiMo V2.6 Flash (Xiaomi): Cost per test (USD) $0.020, Score 76.40%XiGPT 5.5 (OpenAI): Cost per test (USD) $0.74, Score 76.40%OGPT-5.6 Terra (OpenAI): Cost per test (USD) $0.47, Score 77.53%OGemini 3.7 Flash (Google): Cost per test (USD) $1.02, Score 77.53%GGrok 4.6 (xAI): Cost per test (USD) $0.45, Score 78.28%XMuse Spark 1.3 Max (Meta): Cost per test (USD) $0.60, Score 79.03%MGPT-5.6 Luna (OpenAI): Cost per test (USD) $0.050, Score 79.03%OClaude Fable 5 (Anthropic): Cost per test (USD) $1.43, Score 80.52%AKimi K3 (Moonshot): Cost per test (USD) $0.34, Score 80.90%KGemini 3.8 Flash (Google): Cost per test (USD) $1.54, Score 81.27%GClaude Sonnet 5.5 (Anthropic): Cost per test (USD) $0.63, Score 83.15%AGPT-6 Sol (OpenAI): Cost per test (USD) $0.37, Score 83.15%OClaude Opus 5 (Anthropic): Cost per test (USD) $0.89, Score 84.64%AClaude Fable 5.1 (Anthropic): Cost per test (USD) $3.07, Score 85.02%AGPT-5.6 Sol (OpenAI): Cost per test (USD) $1.02, Score 85.77%OGPT-6 Astra (OpenAI): Cost per test (USD) $1.34, Score 87.27%OClaude Opus 5.5 (Anthropic): Cost per test (USD) $0.50, Score 87.64%AClaude Opus 5.5GPT-6 AstraGPT-5.6 Sol
Terminal-Bench 2.1: top 15 by ScoreRanked bars. Leader: Claude Opus 5.5 at 87.64%. Full values in the table below.0%20%40%60%80%100%Claude Opus 5.5 (Anthropic): 87.64%AClaude Opus 5.587.64%GPT-6 Astra (OpenAI): 87.27%OGPT-6 Astra87.27%GPT-5.6 Sol (OpenAI): 85.77%OGPT-5.6 Sol85.77%Claude Fable 5.1 (Anthropic): 85.02%AClaude Fable 5.185.02%Claude Opus 5 (Anthropic): 84.64%AClaude Opus 584.64%GPT-6 Sol (OpenAI): 83.15%OGPT-6 Sol83.15%Claude Sonnet 5.5 (Anthropic): 83.15%AClaude Sonnet 5.583.15%Gemini 3.8 Flash (Google): 81.27%GGemini 3.8 Flash81.27%Kimi K3 (Moonshot): 80.90%KKimi K380.90%Claude Fable 5 (Anthropic): 80.52%AClaude Fable 580.52%GPT-5.6 Luna (OpenAI): 79.03%OGPT-5.6 Luna79.03%Muse Spark 1.3 Max (Meta): 79.03%MMuse Spark 1.3 Max79.03%Grok 4.6 (xAI): 78.28%XGrok 4.678.28%Gemini 3.7 Flash (Google): 77.53%GGemini 3.7 Flash77.53%GPT-5.6 Terra (OpenAI): 77.53%OGPT-5.6 Terra77.53%
Score
Share of terminal tasks completed and verified by tests.
Cost per test (USD)
Average API cost per task, as published.
Terminal-Bench 2.1 scores
RankModelProviderScoreCost per test (USD)HarnessPrice in / out ($/1M tokens)
1 Claude Opus 5.5Anthropic87.64%$0.50Terminus 2$4 / $20
2 GPT-6 AstraOpenAI87.27%$1.34Terminus 2$10 / $50
3 GPT-5.6 SolOpenAI85.77%$1.02Terminus 2$4 / $20
4 Claude Fable 5.1Anthropic85.02%$3.07Terminus 2$10 / $50
5 Claude Opus 5Anthropic84.64%$0.89Terminus 2$5 / $25
6 GPT-6 SolOpenAI83.15%$0.37Terminus 2$2 / $10
7 Claude Sonnet 5.5Anthropic83.15%$0.63Terminus 2$2 / $10
8 Gemini 3.8 FlashGoogle81.27%$1.54Terminus 2$1.5 / $7.5
9 Kimi K3Moonshot80.90%$0.34Terminus 2$3 / $15
10 Claude Fable 5Anthropic80.52%$1.43Terminus 2$10 / $50
Show 15 more rows
RankModelProviderScoreCost per test (USD)HarnessPrice in / out ($/1M tokens)
11 GPT-5.6 LunaOpenAI79.03%$0.050Terminus 2$0.2 / $1.2
12 Muse Spark 1.3 MaxMeta79.03%$0.60Terminus 2$1.25 / $4.25
13 Grok 4.6xAI78.28%$0.45Terminus 2$2 / $6
14 Gemini 3.7 FlashGoogle77.53%$1.02Terminus 2$1.5 / $7.5
15 GPT-5.6 TerraOpenAI77.53%$0.47Terminus 2$2 / $12
16 GPT 5.5OpenAI76.40%$0.74Terminus 2$5 / $30
17 MiMo V2.6 FlashXiaomi76.40%$0.020Terminus 2$0.14 / $0.28
18 DeepSeek V4.1 FlashDeepSeek74.53%$0.10Terminus 2$0.3 / $1.2
19 Claude Sonnet 5Anthropic74.53%$0.53Terminus 2$2 / $10
20 Gemini 3.5 FlashGoogle74.16%$0.87Terminus 2$1.5 / $9
21 Gemini 3.6 FlashGoogle73.78%$1.17Terminus 2$1.5 / $7.5
22 Grok 4.7xAI73.41%$1.08Terminus 2$2 / $6
23 GPT-6 LunaOpenAI73.03%$0.030Terminus 2$0.1 / $0.5
24 Muse Spark 1.3Meta72.28%$0.66Terminus 2$1.25 / $4.25
25 Claude Opus 4.8Anthropic71.91%$2.41Terminus 2$5 / $25

Source: Vals AI leaderboard ↗ · Terminal-Bench project ↗

Top model here: Claude Opus 5.5 (Anthropic). Try prompts for Claude →

ChatJudge does not run these tests. Scores are copied as published by each source, with a link and the date we read them, and only change when we refresh from the source. Different leaderboards use different settings, so compare models within one chart, not across charts. Spotted a mistake? Tell us.