FrontierBench v0.1 (Opus 5 System Card): leaderboard
FrontierBench v0.1, the Terminal-Bench team's successor to Terminal-Bench 2.1: 74 harder containerized command-line tasks emphasizing science and engineering, run with mini-SWE-agent; mean reward over five attempts per task (percent, higher is better). From Table 8.1.A of the Claude Opus 5 System Card (July 2026); Claude 5 runs fell back to Opus 4.8 when safety classifiers refused.
Metric: Mean Reward (%). Source: www-cdn.anthropic.com. Saturation forecast: Around 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 37.5 | #22 (GPT-5.6 Sol) |
| 2 | Claude Opus 4.8 | 18.7 | #39 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=frontierbench-v0-1-opus-5-system-card · How It Works · Data refreshed daily, snapshot 2026-11-05.