Revelo Code Index - Terminal Bench 3.0
Code · 2026-07-29
The Revelo Code Index's Terminal Bench 3.0 task set, called Frontier Bench until July 30, 2026: 20 private tasks across software engineering, business operations, science, machine learning and security in the Terminal-Bench 3.0 format, though none are Terminal-Bench's own tasks. The agent hands in a patch, checkpoint, data artifact or written analysis that a separate verifier container grades against tests and rubrics, and the score is the mean over tasks of three trials.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 31.7 |
| GPT-6 Astra | 30.0 |
| Claude Fable 5 | 25.0 |
| Qwen3.8-Max | 18.3 |
| DeepSeek V4.1 Flash | 16.7 |
| GPT-5.6 Sol | 16.7 |
| Kimi K3 | 15.0 |
| Opus 4.8 | 13.3 |
| GPT-5.6 Luna | 13.3 |
| GPT-5.6 Terra | 10.0 |
| Gemini 3.6 Flash | 8.3 |
| Gemini 3.8 Flash | 8.3 |
| Hy3 | 6.7 |
| DeepSeek-V4-Flash-0731 | 5.0 |
| DeepSeek-V4-Pro | 5.0 |
| GLM-5.2 | 3.3 |
| MiMo-V2.5-Pro | 3.3 |
| MiniMax M3 | 3.3 |
| Grok 4.5 | 1.7 |