What it measures
- Terminal-Bench 4.0
Terminal-Bench 4.0
Terminal-Bench 4.0 result records from the supplied 2026 frontier model reports.
How to read the score
% score; higher is better.
Sourced Ranking
| Rank | Model | Provider | Score | Source | Source Type | Verified |
|---|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5 | Anthropic | 70.6 | Introducing Claude Sonnet 5.5 Accessed 2026-10-02 | official | Yes |
| 2 | Claude Opus 5.5 | Anthropic | 66.4 | Introducing Claude Opus 5.5 Accessed 2026-10-02 | official | Yes |
| 3 | GPT-6 Astra | OpenAI | 57.9 | GPT-6 Astra: A new generation of intelligence Accessed 2026-09-04 | official | Yes |
| 4 | Claude Fable 5.1 | Anthropic | 55.8 | Claude Fable 5.1 Accessed 2026-09-02 | official | Yes |
| 5 | Gemini 3.8 Flash | 19.1 | Gemini 3.8 Flash model card Accessed 2026-09-04 | official | Yes |
Limitations
- Scores may not be directly comparable across model reports, harnesses, or task releases.
- Preserve the exact version, evaluation configuration, and reporting source in each result row.
- Vendor and third-party scores are stored as raw evidence and are not automatically used in the five-axis capability radar.
leaderboard
Terminal-Bench 4.0 source record
Preserve the exact version, evaluation configuration, and reporting source in each result row. Vendor and third-party scores are stored as raw evidence and are not automatically used in the five-axis capability radar.