Terminal-Bench 4.0: leaderboard
Sixty-six containerized terminal tasks weighted toward science-adjacent and frontier engineering work, each agent run five times. Relative to 2.1 it lengthens timeouts and raises RAM and CPU on selected tasks, so a failure is more likely the agent's than the harness's.
Metric: Accuracy (%). Source: www.tbench.ai. Status: years away from saturation. 14 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 | 58.2 |
| 2 | Claude Fable 5.1 | 57.9 |
| 3 | Claude Opus 5 | 51.8 |
| 4 | Claude Fable 5 | 44.5 |
| 5 | GLM-5.3 | 41.8 |
| 6 | GPT-5.6 Sol | 37.3 |
| 7 | Claude Opus 4.8 | 23.6 |
| 8 | GPT-5.6 Terra | 21.5 |
| 9 | Grok 4.6 | 20.3 |
| 10 | Gemini 3.8 Flash | 19.1 |
| 11 | GPT-5.6 Luna | 17.3 |
| 12 | Grok 4.5 | 12.4 |
| 13 | Claude Sonnet 5 | 12.4 |
| 14 | Gemini 3.7 Flash | 11.2 |
Interactive version: theaggregate.ai/benchmark?slug=terminal-bench-4-0 · How It Works · Data refreshed daily, snapshot 2026-09-05.