DrivingBench reports that GPT-6 Astra was the only model to finish its Toyota Corolla course, reaching 100% centerline progress in 5:22 on its second attempt. Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol did not finish their recorded runs. The project announced the test in a post dated September 21, 2026.
How DrivingBench set up the test
The models controlled steering, acceleration and braking in a real Toyota Corolla. DrivingBench says a comma 4 interface handled those vehicle controls over CAN, while the models sent commands through MCP tools in Codex, Cursor or Claude Code.
The test took place in an empty lot. The car’s speed was capped in software at 8 mph, and a human remained in the driver’s seat without steering.
What the progress score measures
DrivingBench defines progress as the distance covered along the course centerline while the car remains within 4 meters of it. The leaderboard also lists GPS-integrated travel distance, a separate measure; the percentages here are centerline progress.
Each model had up to three attempts in one continuous chat. Later attempts followed reflection and continuation prompts, so they built on the same conversation rather than starting fresh. GPT-6 Astra reached 49% on its first attempt, completed the course on its second, and had no third attempt.
Best centerline progress by model
DrivingBench’s leaderboard reports these best progress figures for the four models:
| Model | Best centerline progress |
| GPT-6 Astra | 100% |
| Claude Fable 5.1 | 45% |
| Grok 4.6 | 11% |
| GPT-5.6 Sol | 6% |
What the benchmark says about public-road driving
The result describes a run on a controlled course in an empty lot, with an 8 mph software speed cap and a human in the car. It does not establish how any of the models would perform on public roads.
