GPT-6 Astra completes real car driving course in benchmark test
Original: GPT-6 Astra has gained the ability to drive a car
Why This Matters
First public evidence of a general-purpose LLM completing a real physical driving task end-to-end.
Independent benchmark DrivingBench gave frontier AI models direct control of a Toyota Corolla's steering, accelerator, and brakes on a fixed cone course. GPT-6 Astra (via Codex, medium setting) completed 100% of the course in 5 minutes 22 seconds on its second attempt, while rivals Claude Fable 5.1 and Grok 4.6 failed to finish.
DrivingBench, a research project by Aditya Ramabadran, Simon Mahns, and Tobias Gessler, tests whether large language models can physically operate a real Toyota Corolla around a cone course. Models receive up to three attempts within one continuous chat session and are scored on how far along the course centerline they progress while staying within 4 meters of it.
GPT-6 Astra, run through OpenAI Codex at medium settings, failed on its first attempt at 49% progress, then completed the full course on attempt two — covering 134.7 meters in 5:22, issuing 24 accepted motion commands and consuming 6.6 million tokens at a cost of $7.74.
No other model finished. Claude Fable 5.1 peaked at 45% on its third try. Grok 4.6 topped out at 11%. GPT-5.6 Sol stalled around 6% across all three runs.
Progress is measured along the centerline to the finish zone; a collision freezes progress at whatever point was reached before impact. The benchmark explicitly notes it is research software, not a commercial product.