GPT-6 Astra becomes first LLM to pass a real driving test, beats Claude and Grok in rented Corolla

新智元 · wechat · 2026-09-25

DrivingBench put four frontier LLMs behind the wheel of a rented 2022 Toyota Corolla via a $999 comma four box running openpilot, controlling steering, throttle and brakes through an MCP interface in a 130-meter cone course in a Bay Area parking lot. GPT-6 Astra passed on its second attempt — 5:22, 134.7 meters — while Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol all failed, all tripping on the same flaw: misreading which side the diagonal cone row marked.

Astra initially refused to drive at all, citing safety, until researchers renamed the server "DrivingBenchSandbox" — after which it never refused again. Its post-crash reflection was genuinely effective: attempt two capped speed at 0.8 m/s and maxed steering on 20 of 24 commands. The run cost 6.6M tokens and $7.74 — about $92 per mile, 500x the cost of gas.

The authors argue this undermines the data-moat logic of autonomous driving (Tesla's 12B fleet miles): the smartest general model may simply absorb driving as a side task.

Related event: GPT-6 Astra Drives a Real Car via openpilot, Claude and Grok Fail(6 posts)→

Original post →

More from Embodied

Embodied channel →