Atlas

Benchmarks

← All benchmarks

FeatureArc (first stage)

Code · 2026-10-27

The share of the 73 FeatureArc tasks whose first stage, implemented from the initial repository state, passes both its hidden fail-to-pass and pass-to-pass suites, in percent. It is equivalent to a single-stage feature task, run in the same persistent sessions and budget as the whole-feature chain gate.

Top models (higher is better)

ModelScore
Claude Opus 5.583.6
GPT-6 Astra67.1
GLM-5.364.4
Gemini 3.8 Flash61.6
Grok 4.760.3
Kimi K356.2
DeepSeek V4 Pro 081352.1
Loading Atlas data…