Atlas

Benchmarks

← All benchmarks

RWE-bench

Agents · 2026-10-28

Xplore's private Real World Engineering Bench: 20 engineering design tasks (9 mechanical, 6 aerospace, 5 electrical), such as designing a fatigue-limited diesel crankshaft in CAD, routing a DDR3 memory interface on a populated production board, or meshing and solving turbine-vane CFD, each written by a domain engineer with a reference solution. Agents work in a sandbox with professional CAD and CAE tools for up to 24 hours per run, and a deterministic verifier that runs physics solvers, with no LLM or human grading, passes a design only if it meets every requirement in the brief. Each model runs up to five times per task in its developer's coding agent; the average pass rate is the share of passing runs averaged over the tasks.

Top models (higher is better)

ModelScore
GPT-6 Astra37.0
Claude Opus 5.529.0
GPT-6.1 Sol25.0
Gemini 4 Argon12.0
Claude Fable 5.17.0
Kimi K3.1 Pro1.0
GLM-5.30.0
Grok 4.70.0
Muse Wonder0.0
Loading Atlas data…