BENCHMARK EVIDENCE / ACTIVE
Terminal-Bench
v4.0
agentic coding. Metric: task-success. Direction: higher-is-better. Results below stay separated by evaluator evidence type and comparable group.
3observations
Artificial Analysisevaluator
activestatus
2026-10-04verified
METHOD
Read the score with its configuration.
Scores from different comparable groups are intentionally not merged into one ranking. Reasoning effort, tools, fallback behavior, evaluator and benchmark version stay attached to every observation.
- Terminal-heavy agentic coding tasks. Harness and reasoning configuration materially affect results.
independent
3 observationsartificial-analysis-v4.3.2
| Model | Score | Reasoning | Tools | Source |
|---|---|---|---|---|
| Claude Opus 5.5 | 60.0% | max | benchmark-defined | Evidence ↗ |
| GPT-6 Astra | 54.0% | high | benchmark-defined | Evidence ↗ |
| GPT-6 Sol | 26.0% | high | benchmark-defined | Evidence ↗ |