> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# GPT-4 Just Drove Itself to In-N-Out in a Real Car
- URL: https://wire.fourthweb.ai/gpt-4-just-drove-itself-to-in-n-out-in-a-real-car/
- Published: 2026-10-10T23:00:41.000Z
- Updated: 2026-10-10T23:00:42.000Z
- Description: The moment AI can navigate city streets to grab a burger is the moment we stop pretending these models are just chatbots. Three engineers tested GPT, Claude, and Grok driving a real Toyota Corolla to an In-N-Out, with only one model successfully completing the trip
- Author: Travis Wright
- Tags: AI Agent Economy, DeFi, OpenAI, Anthropic

**The moment AI can navigate city streets to grab a burger is the moment we stop pretending these models are just chatbots.**

### The Summary

- Three engineers tested [GPT, Claude, and Grok driving a real Toyota Corolla](https://www.wired.com/story/ai-is-driving-cars-now-oh-boy/?ref=wire.fourthweb.ai) to an In-N-Out, with only one model successfully completing the trip
- This wasn't simulation—it was a production car on public roads, marking a shift from controlled testing to real-world autonomous decision-making
- The experiment reveals which frontier models can actually reason spatially and sequentially in dynamic environments, not just generate text

### The Signal

The experiment stripped away the usual autonomous vehicle stack. No custom sensors, no proprietary vision systems, no safety drivers hovering over backup controls. Just a stock Toyota Corolla, a camera feed, and three frontier LLMs making every driving decision in real time. [One model made it to the In-N-Out drive-through](https://www.wired.com/story/ai-is-driving-cars-now-oh-boy/?ref=wire.fourthweb.ai). The other two failed within the first mile.

This matters because it proves something the autonomous vehicle industry has been dancing around for years: the hard part isn't perception or prediction. It's reasoning under uncertainty with imperfect information. The same skill that makes a model good at coding or legal analysis.

> "The model that succeeded wasn't the one with the best benchmark scores. It was the one that could hold multiple constraints in working memory while making sequential decisions."

The researchers found three critical differences:

- Spatial reasoning: tracking position relative to lanes, other cars, and navigation targets simultaneously
- Temporal planning: knowing when to brake early versus coast versus accelerate based on upcoming intersections
- Constraint juggling: following traffic rules while optimizing for speed and passenger comfort

[Claude](https://wire.fourthweb.ai/tag/anthropic/) drove conservatively, treating every yellow light like a stop sign and slowing to 15 mph before every turn. GPT hallucinated a pedestrian crossing and emergency-braked on an empty road. Grok completed the 4.2-mile route in 11 minutes, including the drive-through order, without human intervention.

The engineers published their full methodology and failure logs. They're not selling a product. They're showing that general-purpose reasoning models can control physical systems in unstructured environments without years of specialized training data. That's the unlock. If an LLM can drive, it can operate construction equipment, pilot drones, or control robotic arms in warehouses.

### The Implication

This isn't about self-driving cars replacing Waymo. It's about agentic systems controlling physical infrastructure with the same models already running backend automation and customer service. The next wave of Web4 agents won't just live in APIs and databases. They'll move atoms, not just bits.

Watch for two things: hardware companies bolting LLM-based control systems onto existing equipment, and safety regulators scrambling to define liability when a general-purpose model makes a physical mistake. The distance from "AI can drive to In-N-Out" to "AI can operate the forklift fleet overnight" is shorter than most people think.

### Sources

[Wired AI](https://www.wired.com/story/ai-is-driving-cars-now-oh-boy/?ref=wire.fourthweb.ai)