GPT-6 Astra successfully navigates a track in a real car
GPT-6 Astra successfully navigates a track in a real car
The DrivingBench team tested four AI models on controlling a real Toyota Corolla. Only GPT-6 Astra managed to complete the track.
The DrivingBench team connected four AI models to the steering wheel, gas, and brake pedals of a 2022 Toyota Corolla. The car moved along a cone-marked track in an empty parking lot, while the models received images from cameras and issued commands through control tools.
Astra turned out to be the only model that reached the finish line. In the first attempt, it completed 49% of the route, then analyzed its mistake and finished the track in 5 minutes and 22 seconds on the second try. The best result for Claude Fable 5.1 was 45%; Grok 4.6 and GPT-5.6 Sol failed to pass the first turn.
A human was behind the wheel at all times, ready to hit the brakes. The authors conducted only one series of attempts for each model, so the result does not indicate road reliability. However, the experiment already shows that a general-purpose model is capable of controlling a physical car and correcting its actions during the task.
Why it matters
AnalysisThe experiment confirms that modern multimodal general-purpose AI models can effectively control complex physical systems in real-time, adapting to errors as they occur.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
GPT-6 Astra: Guide to OpenAI's New Model
Comprehensive overview of OpenAI's flagship GPT-6 Astra: agentic architecture, Computer Use capabilities, software engineering in Codex, cybersecurity benchmarks, prompt engineering tips, and API migration guide.
How to Create Realistic AI Lip Sync: Script, Emotions, Voice, and Facial Synchronization
AI lip sync is everywhere: product demos, training materials, social media, multilingual ad campaigns. At its core lies the same technology: making a digital face speak convincingl...
Hallucination Risks in Real-World Tasks
An analysis of the practical consequences and legal risks arising from blind trust in fabricated facts and false sources generated by language models in financial reports, legal claims, and medical advice.
Web Search in Chat (Real-Time Internet Search)
An integrated web search feature within ChatGPT, Claude, and Gemini (globe icon). It enables the model to retrieve fresh data from websites for the current day and provide answers with direct links to sources.