AI models tested on flying fighter jets in Top Gun canyon simulation
A user tested several frontier LLMs to see if they could navigate a complex flight path, revealing significant gaps in their spatial reasoning and safety.
The short version
- Models were given 56 raw data points including height, speed, and G-force to navigate a canyon.
- Claude Opus 5.5 refused to pilot the aircraft, while others like Gemini Carbon failed the first gate.
- ChatGPT Astra lasted the longest but eventually crashed after misjudging distance under a bridge.
- API costs for the flights ranged from $0.00 for Opus to $14.57 for Claude Fable 5.2.
Why it matters
This test demonstrates that while LLMs are improving at coding, they still struggle with real-time spatial awareness and complex physics. It serves as a practical reminder that AI is not yet ready for high-stakes, real-world physical navigation.
The Top Gun Challenge
A user on the ClaudeAI subreddit recently designed a high-stakes simulation to test the spatial reasoning of frontier large language models (LLMs). The goal was to see if these AI systems could navigate a Boeing F/A-18 Super Hornet through the iconic canyon scene from the movie Top Gun: Maverick.
The experiment pitted several prominent models against each other, including ChatGPT Astra, Google Gemini Carbon, Claude Fable 5.2, and Claude Opus 5.5. To make the test fair, the human actor playing Maverick flew the plane as he did in the film, while the AI models had to figure out the flight path entirely on their own.
This setup provides a more tangible way to measure AI capabilities than standard benchmarks. Instead of just answering questions, the models had to translate raw data into physical movements in a three-dimensional environment. It highlights the difficulty of moving from text generation to real-world physics.
How the simulation worked
The technical framework for the test relied on the models writing their own autopilot code. Every 10 seconds, the simulation provided the AI with the jet's full state. This included 56 raw numbers representing variables such as height, speed, G-force, clearance ahead, and the location of the next gate.
The models were not given any flying tips or pre-written instructions. They had to interpret these numbers and set the controls accordingly. To allow for improvement, each model was given 10 test flights to debug its code and receive feedback on its performance before the final run.
This iterative process mimics how software engineers use AI to build complex systems. By giving the models multiple attempts, the user could see if the AI was capable of learning from its mistakes or if it would consistently fail the same maneuvers in the canyon.
Performance and failures
The results showed a wide variety of behaviors among the different models. Claude Opus 5.5 notably refused to pilot the machine of war, which the user attributed to the model's internal safety guardrails regarding dangerous activities. This highlights how safety training can sometimes prevent an AI from completing a specific task.
Other models struggled with the basic physics of the environment. Both Fable 5.2 and Gemini Carbon were eliminated almost immediately because they missed the very first gate. These failures suggest that the models may lack the precise spatial awareness required for tight maneuvers.
ChatGPT Astra was the most persistent of the group, holding its own against the human pilot for an impressive amount of time. However, it eventually crashed after misjudging the distance while flying under a bridge. This indicates that while Astra could handle some complexity, it still lacked the reliability needed for safe navigation.
The cost of AI flight
Beyond the performance of the models, the experiment also tracked the financial cost of running these complex simulations via APIs. The costs varied significantly depending on the model used. For instance, Claude Opus 5.5 cost $0.00 to run, likely due to its refusal to participate in the flight.
The other models required significant computational resources to process the 56 data points every 10 seconds. ChatGPT Astra cost $1.92 per flight, while Google Gemini Carbon cost $2.27. These figures provide a glimpse into the ongoing expenses associated with running high-frequency inference tasks.
Claude Fable 5.2 was the most expensive participant, costing $14.57 for a single flight. This disparity in pricing shows how different architectures and token usage patterns can lead to vastly different costs for the same task. It serves as a reminder that
Limitations of current AI
The experiment serves as a stark reminder that current LLMs are not yet ready for high-stakes physical tasks. While they can write code and process data, they still struggle with the nuances of real-time navigation and physical consequences. The fact that most models crashed or refused to fly suggests a long way to go for autonomous flight.
The user noted that these types of benchmarks are often more informative than the official charts released by AI companies. They provide a
What's next for navigation
Moving forward, the focus for developers will likely shift toward improving spatial reasoning and
What to watch
- How models will handle more complex physics in future simulations.
- Whether safety guardrails will continue to prevent models from performing 'dangerous' tasks like piloting aircraft.
Try it
Look up the ClaudeAI subreddit to see other user-generated benchmarks for frontier models.
Sources and links
First reported by xda-developers.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.