These Researchers Made AI Drive a Toyota Corolla to Get In-N-Out

Three Axiom engineers used OpenAI’s GPT-6 Astra to steer a 2024 Toyota Corolla through an In-N-Out drive-thru, with a safety driver ready to brake. In their DrivingBench parking-lot test, Astra completed the course slowly, Claude Fable 5.1 covered 45%, and Grok 11%; the article presents physical reasoning as an emerging capability, while emphasizing that general-purpose models remain far from reliable drivers and could pose risks.

Analysis

Physical AI: capability signal, not deployment signal. The market may overread a model’s ability to improvise in a constrained demo as evidence that general-purpose models can replace purpose-built autonomy. The economic bottleneck is not just scene understanding: it is reliable control across rare edge cases, safety validation, and liability. That leaves a near-term advantage with operators that have accumulated driving data and built dedicated safety systems, including Alphabet’s Waymo; the demo does not establish a comparable change in Tesla’s or Toyota’s economics.

Over 1–3 months, watch for independent, repeatable benchmark results and evidence that performance transfers beyond curated routes. If progress is real, the first beneficiaries may be robotics and autonomy developers able to use general models as a reasoning layer, while specialized software and data providers face eventual pricing pressure. Over 6–18 months, broader multimodal capability could lower the cost of building robot applications—but hardware integration, deployment reliability, and regulatory approval remain gating items.

The contrarian read is that the demo is more revealing about model adaptability than about safe autonomy: a cautious, supervised drive is not a commercially deployable system. A serious incident involving general-purpose models controlling vehicles or robots could trigger scrutiny that weighs on the whole physical-AI theme. Treat claims from model developers and benchmark creators as provisional until independently reproduced.

Key Decisions for Investors

• No event-driven position in GOOG, TSLA, or TM on this evidence alone; the demo provides no verified change to revenue, margins, or deployment timelines.
• Avoid extrapolating physical-AI demonstrations into near-term autonomous-driving revenue. Reassess only after independent testing shows robust performance across varied routes and without a safety driver, alongside a credible regulatory path.
• Set an alert for verified deployment milestones, safety incidents, or regulatory restrictions. A broad incident-driven pullback could create a better entry in established autonomy exposure; repeated, transferable benchmark gains would challenge the view that specialized stacks retain an advantage.
• Falsify the cautious thesis if independent tests demonstrate reliable real-world control across edge cases and operators disclose commercially meaningful deployments or unit economics; conversely, a material safety event would strengthen the regulatory-risk case.