September 23, 2026
The AI coding agent takes the wheel: GPT-6 Astra covered 134,7 m
On 24 September, GPT-6 Astra in Codex completed its second DrivingBench run: 100% of the track, 134,7 m in 5:22. It used 2,6 million tokens and $7,74. The first attempt reached 49%, or 67,3 m, and cost $2,01.
On its first attempt, GPT-6 Astra stopped halfway through the track. On the second, it took the car to the finish, while GPT-5.6 Sol scored 6% in each of three runs, Claude Fable 5.1 reached 45%, and Grok 4.6 reached 11%.
The agent has three tools. DrivingBench gives Codex the MCP tools observe, set_motion, and stop_now. Through set_motion, the model sets direction, steering, speed from 0,5 to 3,5 m/s, and command duration.
The testbed is open. The harness is available on GitHub: its authors suggest installing dependencies with `uv sync --locked`, then connecting chat apps with `uv run drivingbench install --client codex --client claude --client cursor --host <comma>` and restarting them. You need comma four with openpilot v0.11.2, Toyota TSS2, macOS with Python 3.12, an empty private lot, and cones.
After installation, the testbed opens locally, collects parameters in Bring online, and generates a chat prompt. The authors allow runs only on an empty private lot with a driver at the wheel and a foot on the brake.
The protocol gives the model up to three attempts in one chat, and after a failure adds a request to analyze the error and continue to the context.
