GPT-6 Astra Only Model to Finish DrivingBench Cone Course, in Just 2 Tries
FlorianGallwitz · x · 2026-09-23
DrivingBench published full results from its cone-course driving benchmark, where each model gets 3 attempts in the same continuous chat to test in-context learning.
Per video shared by aramabadran, GPT-6 Astra was the only model to complete the entire course, finishing in just 2 tries — all other models failed to run the full course. It's a rare public eval of an LLM directly controlling driving.
Related event: GPT-6 Astra Is the Only Model to Pass Autonomous Driving Course(2 posts)→
More from Models
- Browser Use Bench v2: GPT-6 Sol scores 66.9, beating Opus 5.5 at 3.5x lower cost — airesearch12 · 2026-09-23
- Early Access Claims: Claude Opus 5.5 Builds Dark Souls-Style Games, Reportedly Far Beyond Expectations — FinanceYF5 · 2026-09-23
- $50k bounty unclaimed: no one has found pre-2022 French literature flagged 100% AI — Afinetheorem · 2026-09-23
- Pangram defends its AI detector: false positives measured on test sets, never on full book chapters — cephaloform · 2026-09-23
- GPT-6 Sol reportedly worse than 5.6 Sol on DeepSWE and computer use; mocked as rebranded Terra — kristoph · 2026-09-23
- Mini benchmark: Opus 5.5 draws better but burns 10x more tokens than Astra — OfirPress · 2026-09-23