Terminal-Bench 4 · System reference
What it measures and how
Fix Terminal-Bench 4.0 and retain model, Agent version, reasoning tier and run count separately. Different tool systems do not represent common model conditions.
How this evidence is used
Raw results for reference: automatically collected after locking the official submission list and data version. Retain each system configuration without selecting a representative model score. Cross-checks only; excluded from overall and coding ranks.
Benchmark results
These are system scores for models paired with different Agents under different run conditions, for reference only. Up to 30 configurations are shown per benchmark. Anonymous test models are omitted.
| Source rank | Source model name | Raw result | Run configuration |
|---|---|---|---|
| 1 | claude-opus-5-5Anthropic | 64.8% | Claude Code 2.1.284 · max reasoning |
| 2 | claude-sonnet-5-5Anthropic | 61.8% | Claude Code 2.1.284 · max reasoning |
| 3 | gpt-6.1-solOpenAI | 58.2% | Codex 0.159.0 · max reasoning |
| 4 | claude-fable-5-1Anthropic | 57.9% | Claude Code 2.1.257 · max reasoning |
| 5 | claude-opus-5Anthropic | 51.8% | Claude Code 2.1.231 · max reasoning |
| 6 | gpt-6-solOpenAI | 49.4% | Codex 0.159.0 · max reasoning |
| 7 | claude-fable-5Anthropic | 44.5% | Claude Code 2.1.231 · max reasoning |
| 8 | glm-5.3Z.ai | 41.8% | Claude Code 2.1.207 · max reasoning |
| 9 | gpt-5.6-solOpenAI | 37.3% | Codex 0.149.1 · max reasoning |
| 10 | glm-5.3-flashZ.ai | 35.8% | Claude Code 2.1.285 · Reasoning tier not reported by the source |
| 11 | qwen3.8-max-0902Alibaba | 27.0% | Claude Code 2.1.285 · max reasoning |
| 12 | claude-opus-4-8Anthropic | 23.6% | Claude Code 2.1.231 · max reasoning |
| 13 | gpt-5.6-terraOpenAI | 21.5% | Codex 0.149.1 · max reasoning |
| 14 | grok-4.6xAI | 20.3% | Grok Build 1.0.5 · Reasoning tier not reported by the source |
| 15 | gemini-3.8-flashGoogle | 19.1% | mini-SWE-agent 2.4.6 · high reasoning |
| 16 | gpt-5.6-lunaOpenAI | 17.3% | Codex 0.149.1 · max reasoning |
| 17 | gpt-6-lunaOpenAI | 16.4% | Codex 0.159.0 · max reasoning |
| 18 | muse-spark-1.3Meta | 14.6% | Muse Code 1.4.1 · xhigh reasoning |
| 19 | claude-sonnet-5Anthropic | 12.4% | Claude Code 2.1.231 · max reasoning |
| 19 | grok-4.5xAI | 12.4% | Grok Build 1.0.5 · Reasoning tier not reported by the source |
| 21 | gemini-3.7-flashGoogle | 11.2% | mini-SWE-agent 2.4.6 · high reasoning |
Limitations and attribution
Models use different tools such as Claude Code, Codex, GrokBuild and mini-SWE. AA v4.3 already includes Terminal-Bench v4, so it must not add overall vote weight again.
Data license: Apache 2.0 · Official result submissions in harbor-framework/terminal-bench; retain attribution, protocol and change notes.
Results are published by Harbor / Terminal-Bench. Raw scores and the e/acc consensus index use different scales and cannot be added directly.