RWE-bench Pass@5
Agents · 2026-10-28
Share of RWE-bench's 20 private engineering design tasks that a model passes in at least one of its five 24-hour runs, from the same runs as the RWE-bench average pass rate. A run passes only when Xplore's deterministic verifier, which runs physics solvers, finds that the design meets every requirement.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 70.0 |
| GPT-6 Astra | 60.0 |
| GPT-6.1 Sol | 50.0 |
| Gemini 4 Argon | 35.0 |
| Claude Fable 5.1 | 20.0 |
| Kimi K3.1 Pro | 5.0 |
| GLM-5.3 | 0.0 |
| Grok 4.7 | 0.0 |
| Muse Wonder | 0.0 |