EVALUATIONS / AGENTS
Agent benchmarks.
Model + harness, tested together.
Terminal-Bench 4.0
Collected submissions · ranked within this benchmark version.
| Rank | Model + harness | Pass rate | Dates + evidence |
|---|---|---|---|
| Rank 1 | Opus 5.5anthropic/claude-opus-5-5Harnessclaude-code 2.1.284Reasoning effortmax Model profile ↗ | Pass rate64.85% |
|
| Rank 2 | Sonnet 5.5anthropic/claude-sonnet-5-5Harnessclaude-code 2.1.284Reasoning effortmax Model profile ↗ | Pass rate61.82% |
|
| Rank 3 | GPT-6.1 Solopenai/gpt-6.1-solHarnesscodex 0.159.0Reasoning effortmax Model profile ↗ | Pass rate58.18% |
|
| Rank 4 | Fable 5.1anthropic/claude-fable-5-1Harnessclaude-code 2.1.257Reasoning effortmax Model profile ↗ | Pass rate57.88% |
|
| Rank 5 | Opus 5anthropic/claude-opus-5Harnessclaude-code 2.1.231Reasoning effortmax Model profile ↗ | Pass rate51.82% |
|
| Rank 6 | GPT-6 Solopenai/gpt-6-solHarnesscodex 0.159.0Reasoning effortmax Model profile ↗ | Pass rate49.39% |
|
| Rank 7 | Fable 5anthropic/claude-fable-5Harnessclaude-code 2.1.231Reasoning effortmax Model profile ↗ | Pass rate44.55% |
|
| Rank 8 | GLM-5.3anthropic/glm-5.3Harnessclaude-code 2.1.207Reasoning effortmax Model profile ↗ | Pass rate41.82% |
|
| Rank 9 | GPT-5.6 Solopenai/gpt-5.6-solHarnesscodex 0.149.1Reasoning effortmax Model profile ↗ | Pass rate37.27% |
|
| Rank 10 | GLM-5.3-Flashz-ai/glm-5.3-flashHarnessclaude-code 2.1.285Reasoning effortnone Model profile ↗ | Pass rate35.76% |
|
| Rank 11 | Qwen3.8-Max-0902qwen/qwen3.8-max-0902Harnessclaude-code 2.1.285Reasoning effortmax Model profile ↗ | Pass rate26.97% |
|
| Rank 12 | Opus 4.8anthropic/claude-opus-4-8Harnessclaude-code 2.1.231Reasoning effortmax Model profile ↗ | Pass rate23.64% |
|
| Rank 13 | GPT-5.6 Terraopenai/gpt-5.6-terraHarnesscodex 0.149.1Reasoning effortmax Model profile ↗ | Pass rate21.52% |
|
| Rank 14 | Grok 4.6xai/grok-4.6Harnessgrok-build 1.0.5Reasoning effortnone Model profile ↗ | Pass rate20.3% |
|
| Rank 15 | Gemini 3.8 Flashgemini/gemini-3.8-flashHarnessmini-swe-agent 2.4.6Reasoning efforthigh Model profile ↗ | Pass rate19.09% |
|
| Rank 16 | GPT-5.6 Lunaopenai/gpt-5.6-lunaHarnesscodex 0.149.1Reasoning effortmax Model profile ↗ | Pass rate17.27% |
|
| Rank 17 | GPT-6 Lunaopenai/gpt-6-lunaHarnesscodex 0.159.0Reasoning effortmax Model profile ↗ | Pass rate16.36% |
|
| Rank 18 | Muse Spark 1.3meta/muse-spark-1.3Harnessmuse-code 1.4.1Reasoning effortxhigh Model profile ↗ | Pass rate14.55% |
|
| Rank 19 | Grok 4.5xai/grok-4.5Harnessgrok-build 1.0.5Reasoning effortnone Model profile ↗ | Pass rate12.42% |
|
| Rank 19 | Sonnet 5anthropic/claude-sonnet-5Harnessclaude-code 2.1.231Reasoning effortmax Model profile ↗ | Pass rate12.42% |
|
| Rank 21 | Gemini 3.7 Flashgemini/gemini-3.7-flashHarnessmini-swe-agent 2.4.6Reasoning efforthigh Model profile ↗ | Pass rate11.21% |
|
Each result belongs to the full model, harness version and reasoning configuration shown. Confidence intervals describe the source’s reported uncertainty; tied pass rates share a rank. “n” is the reported sample count. Release and check dates are separate from evaluation dates.