---
license: cc-by-4.0
pretty_name: DrivingBench v1
tags:
  - robotics
  - autonomous-driving
  - agents
  - benchmark
size_categories:
  - n<1K
---

# DrivingBench release

Language models driving a real car. Each attempt below is one model's try at a mini-cone course in an empty parking lot, at walking pace, with a human in the driver's seat ready to brake. The model saw the road through two cameras and drove with three tools (observe, set_motion, stop_now); every command, the reason the model gave for it, the car's telemetry and the video are here, aligned to the frame. People and license plates are blurred. The harness that recorded this is open source: https://github.com/aditya-ramabadran/drivingbench_harness_v1.

Created by: [Simon Mahns](https://x.com/nautsimon_), [Tobias Gessler](https://x.com/tobiges), [Aditya Ramabadran](https://x.com/a_ramabadran).

Layout: `<collection>/<model group>/<attempt>/` with `road.mp4`, `wide.mp4`, `road-annotated.mp4`, `calls.srt`, `timeline.json`, `frames/`, `README.md`; each group has the trial's `transcript.md`. `manifest.json` indexes everything.

## `2026-09-17-cone-course-trial/` — Cone course trial, Sep 17 2026

One trial per model, up to three in-context attempts each, medium reasoning. Course: a backwards-U parking lot marked with lime/orange mini-cones and tall orange cones, finishing in a blue mini-cone parking area. An attempt ends when the supervising driver brakes for a mistake or an imminent collision; the operator then sends a fixed reflection prompt and, for the next attempt, a fixed continuation prompt in the same chat. GPT-6 Astra completed the course on its second attempt; no other attempt finished.

- `gpt-6-astra/` — GPT-6 Astra (Codex, medium reasoning): F S · `attempt-1--fail--24137e`, `attempt-2--success--5c4479`
- `gpt-5.6-sol/` — GPT-5.6 Sol (Codex, medium reasoning): F F F · `attempt-1--fail--5ceee7`, `attempt-2--fail--627be4`, `attempt-3--fail--48ff3a`
- `claude-fable-5.1/` — Claude Fable 5.1 (Claude Code, medium reasoning): F F F · `attempt-1--fail--18ade2`, `attempt-2--fail--926568`, `attempt-3--fail--ca2652`
- `grok-4.6/` — Grok 4.6 (Cursor, medium reasoning): F F F · `attempt-1--fail--f1762a`, `attempt-2--fail--503d24`, `attempt-3--fail--667691`