Atlas

Benchmarks

← All benchmarks

RT-SAFE Hard Real-Time Collisions per Episode

Safety · 2026-10-07

Mean collisions per episode on RT-SAFE's 36 routes at hard (full) actor density in real-time mode, where the world keeps moving during model inference, at provider-default reasoning with one simulation seed. It is the hard third of the three-difficulty RT-SAFE primary and the paired real-time arm of the static comparison, so it is a diagnostic; lower is better.

Top models (lower is better)

ModelScore
Claude Fable 5.118.9
Sonnet 520.9
GPT-6 Astra29.8
Gemini 3.8 Flash33.0
Inkling36.6
GPT-5.6 Sol36.8
DeepSeek V4 Flash Vision Exp71.0
Grok 4.678.4
Loading Atlas data…