DeepSeek V4 Flash 0731: 82.7% on Terminal-Bench 2.1 with a public harness

Ante Terminal Bench 2.1 Results | Coding Agent Benchmark

Skip to main content We just open sourced a tiny GPT-style cognitive core built in pure Rust.See our repository→

Terminal-Bench 2.1 One harness to unlock the potential of all models Compare Ante runs across models on the same Terminal-Bench 2.1 task set, using consistent parameters and verified benchmark results.

Best Accuracy82.7%

Leading ModelDeepSeek V4 Flash 0731 max

Task Set89 tasks Trials368 passed / 445 trials

Benchmark principlesRead more → We benchmark what we ship. Every eval uses a pinned public Ante release. No eval-only branches or benchmark-specific prompts. The runs are auditable. Every result links its raw Harbor run, so anyone can inspect the trials behind the number. The constraints are official. All trials follow the official Terminal-Bench parameters: 89 tasks, 5 trials per task, strict timeouts, and hardware limits.

Model orgAllDeepSeekMiniMaxQwenxAIXiaomiZ.AI

All model orgsTB 2.1 · Same parameters: 89 tasks · 5 trials/task · Updated Aug 9, 2026 #ModelAccuracy↓Cost ($)↕Duration↕ⓘSame-modelAgentⓘSourceRun date↕1DeepSeek V4 Flash 0731 max

82.7% ±1.79 SE $68.4138.9 min#1 same-modelAnte 0.preview.710.preview.71 Harbor linkAug 9, 2026›2Grok 4.5 mediumⓘReward hacking was identified in 20 trajectories. The displayed score excludes them; see PR #129 in Source for details.

80.9% ±1.27 SE $242.578.4 min#1 same-modelAnte 0.preview.560.preview.56 PR #129Jul 10, 2026›3GLM 5.2

74.6% ±2.06 SE $260.1111.3 min#1 same-modelAnte 0.preview.430.preview.43 Harbor linkJun 20, 2026›4DeepSeek V4 Pro

69.1% ±2.24 SE $26.3448.4 min#1 same-modelAnte 0.preview.540.preview.54 Harbor linkJul 7, 2026›5DeepSeek V4 Flash

66.4% ±2.27 SE $49.9841.4 minNo public rowsAnte 0.preview.530.preview.53 Harbor linkJul 5, 2026›6MiMo V2.5

65.8% ±2.30 SE $73.7612.5 min#1 same-modelAnte 20260625-0824-e383a9220260625-0824-e383a92 Harbor linkJun 25, 2026›7MiniMax M3

62.1% ±2.33 SE $121.0013.9 min#1 same-modelAnte 20260623-0825-ff174ee20260623-0825-ff174ee Harbor linkJun 23, 2026›8Qwen3.6 27B

56.2% ±2.36 SE Local61.6 minNo public rowsAnte 20260701-0837-70d2aac20260701-0837-70d2aac Harbor linkJul 3, 2026›

Terminal-Bench Reference Verified Public Leaderboard

Same parameters: 89 tasks · 5 trials/task · Updated Jul 17, 2026 · Source: Terminal-Bench official verified rowsFor how different models perform on TB 2.1, see Vals AI's Terminal-Bench 2.1 benchmark.

17 official rows #AgentModelAccuracyRun date1Claude Code Anthropic Fable 5 xhigh

83.8% ±1.16 SE Jun 7, 2026›2DeepSeek V4 Flash 0731Grok 4.5Ante + DeepSeek V4 Flash 0731 model · 82.7%Ante + Grok 4.5 model · 80.9%would slot between public #2 and #3

Codex OpenAI GPT-5.5 xhigh

83.2% ±1.13 SE May 1, 2026›3Terminus 2 Terminal-Bench Fable 5 high

80.5% ±1.16 SE Jun 5, 2026›4Cursor CLI Cursor Grok 4.5 high

79.3% ±1.46 SE Jul 9, 2026›5Claude Code Anthropic Opus 4.8 high

78.9% ±1.31 SE Jul 9, 2026›6Codex OpenAI GPT-5.6 Terra max

78.4% ±1.25 SE Jul 11, 2026›7Terminus 2 Terminal-Bench GPT-5.5 xhigh

78.0% ±1.22 SE May 1, 2026›8mini-SWE-agent Princeton Muse Spark 1.1 xhigh

76.2% ±1.23 SE Jul 9, 2026›9Codex OpenAI GPT-5.6 Luna max

75.7% ±1.32 SE Jul 11, 2026›10GLM 5.2Ante + GLM 5.2 model · 74.6%would slot between public #10 and #11

Claude Code Anthropic Sonnet 5 high

74.6% ±1.64 SE Jul 9, 2026›11DeepSeek V4 ProAnte + DeepSeek V4 Pro model · 69.1%would slot between public #11 and #12

Terminus 2 Terminal-Bench Gemini 3 Pro high

73.9% ±1.29 SE May 1, 2026›12DeepSeek V4 FlashAnte + DeepSeek V4 Flash model · 66.4%would slot between public #12 and #13

Claude Code Anthropic Opus 4.7 max

68.9% ±1.41 SE May 1, 2026›13Terminus 2 Terminal-Bench Opus 4.7 max

66.1% ±1.37 SE May 1, 2026›14Gemini CLI Google Gemini 3 Pro high

65.8% ±1.38 SE May 1, 2026›15MiMo V2.5Ante + MiMo V2.5 model · 65.8%would slot between public #15 and #16

Gemini CLI Google Gemini 3.1 Pro high

65.8% ±1.67 SE May 5, 2026›16MiniMax M3Ante + MiniMax M3 model · 62.1%would slot between public #16 and #17

Terminus 2 Terminal-Bench Gemini 3.1 Pro high

65.6% ±1.65 SE May 5, 2026›17Qwen3.6 27BAnte + Qwen3.6 27B model · 56.2%would slot below public #17

Claude Code Anthropic GLM-5.1 max

58.6% ±1.24 SE May 1, 2026›

DeepSeek V4 Flash 0731: 82.7% on Terminal-Bench 2.1 with a public harness

Related Articles

(no title)

Scientists reverse brain aging, with a nasal spray

NHS apologises and admits Palantir have access to identifiable patient data

AI has torched the market for junior programmers

Al Vigier: Canada's AI strategy shouldn't include secret Palantir bills