GPT-6 Astra Only Model to Complete DrivingBench Course in Real Toyota Corolla

DrivingBench lets frontier models take control of a real Toyota Corolla's steering, throttle, and brakes on a fixed cone course, testing sustained driving and planning over long-horizon tasks. Each model gets up to three attempts within one continuous conversation to test in-context learning. GPT-6 Astra was the only model to complete the full course: it finished just 49% of the course on its first attempt, then ran the entire course on its second attempt after being asked to reflect on its mistakes.

Confirmed

• Per data shared by @petrusenkomax, GPT-6 Astra's second attempt covered 134.7 meters in 5 minutes 22 seconds, consuming 6 million tokens at a cost of $7.74.
• @1411337 and Polymarket both relayed that GPT-6 Astra was the only model to finish.
• Full model-by-model results were published by DrivingBench, with original video from aramabadran.

Unconfirmed

• Polymarket itself noted the result is currently a single-source relay with no full test report attached.
• The specific list of competing models and their detailed scores were not given in the posts.

Why it matters

The test demonstrates the effectiveness of a reflect-and-retry loop in long-horizon physical-world tasks: learning from a first failure within the same conversation and then completing the course is a landmark case for embodied AI evaluation.

2026-09-23 ~ 2026-09-24 · 5 related posts

• [source] GPT-6 Astra Is the Only Model to Pass Autonomous Driving Course, on Attempt 2 — 141_1337 · 2026-09-23
• GPT-6 Astra Only Model to Finish DrivingBench Cone Course, in Just 2 Tries — FlorianGallwitz · 2026-09-23
• GPT-6 Astra drives a real Toyota Corolla through an obstacle course, the only AI to finish DrivingBench — Polymarket · 2026-09-23
• [source] GPT-6 Astra only model to finish DrivingBench: 134.7m in 5:22 for $7.74 — petrusenko_max · 2026-09-24
• GPT-6 Astra completes entire DrivingBench course on second try after self-reflection — cedric_chee · 2026-09-24