Driving-Bench: Language Models Steer a Real Toyota
Four chatbots with no driving training took the wheel of a Toyota Corolla in an Axiom experiment, and only one finished the course.
In short
Only GPT-6 Astra drove the test course in the Toyota Corolla to 100 percent, while Claude Fable 5.1 managed 45 percent, Grok 4.6 reached 11 percent and GPT-5.6 Sol 6 percent.
At a glance
- Three developers at the startup Axiom had four language models steer a Toyota Corolla through a test course.
- GPT-6 Astra covered 100 percent of the course, driving slowly but never stopping.
- Claude Fable 5.1 reached 45 percent, Grok 4.6 managed 11 percent and GPT-5.6 Sol 6 percent.
- The rig is published on GitHub as Driving-Bench and needs the code plus extra tools and cameras.
- No model was trained to drive; the source gives no course length, speed or safety precautions.
One of four general-purpose language models finished the test course in a Toyota Corolla: GPT-6 Astra completed 100 percent of it. Claude Fable 5.1 got to 45 percent, Grok 4.6 to 11 percent and GPT-5.6 Sol to 6 percent. None of them had been trained to drive.
Chatbots behind the wheel of a stock car
Self-driving cars normally run on algorithms built and trained for that single job. Three developers at the startup Axiom took the opposite route, according to German tech outlet t3n: they wired up general-purpose models from OpenAI, Anthropic and X.ai, tools designed for text, code and images. All four got the car moving. Only one got it to the finish.
A lopsided scoreboard
- GPT-6 Astra: 100 percent of the course, slow but uninterrupted
- Claude Fable 5.1: 45 percent
- Grok 4.6: 11 percent
- GPT-5.6 Sol: 6 percent
The gap matters more than the ranking. Fifty-five percentage points separate first place from second, which is wider than the spread across the remaining three models. Inside a single vendor's lineup the jump is just as stark: GPT-6 Astra scored sixteen times what GPT-5.6 Sol did.
The rig is on GitHub
Axiom released its setup as Driving-Bench on GitHub. With the code, some additional tooling and cameras, other researchers can run the experiment themselves. That is the part worth watching: a published rig can be contradicted by someone else's run, which an in-house demo video cannot.
Why a text model can steer at all
The team points at the training data. Alongside images and video, 3D models now feed into large language model training, which may leave something useful behind for spatial tasks. One developer joked that maybe AGI is already here. That is a punchline, not a measurement — what is documented here is distance covered on a closed course.
What the report does not establish
The available source names neither the three developers nor the layout, length or obstacles of the course. Driving speed, the time each model took per decision and the safety setup are all absent as well. Without them there is no way to judge how hard the task was, or how far these percentages sit from driving in traffic. The full write-up is at t3n.
FAQ
Which AI model won the driving test?
GPT-6 Astra. It was the only model to complete 100 percent of the test course in the Toyota Corolla, although it drove slowly.
How did Claude Fable 5.1 do in the driving test?
Claude Fable 5.1 covered 45 percent of the course, placing second ahead of Grok 4.6 at 11 percent and GPT-5.6 Sol at 6 percent.
Can I reproduce the experiment with Driving-Bench?
The setup is published on GitHub as Driving-Bench and reportedly needs the code plus additional tools and cameras. The source says nothing about cost or required safety clearances.
