The rapid evolution of artificial intelligence has introduced systems capable of writing complex code, analyzing massive datasets, and generating nuanced text in seconds. Yet, despite these incredible capabilities, a lingering skepticism remains regarding how these models would handle real-world physical coordination and high-stakes split-second decision-making. Would anyone truly trust a large language model to pilot a commercial airliner carrying hundreds of passengers? For most people, the answer remains a firm no, and a recent community-driven experiment offers a rather vivid, highly entertaining justification for that hesitation.

Watching some of the most powerful and advanced large language models in the world attempt to fly fighter jets—and promptly nosedive straight into the ground—serves as a stark reminder that artificial intelligence still has significant limitations when it comes to translating digital reasoning into physical mastery. Far from a sterile laboratory benchmark, this latest test brought the theoretical capabilities of frontier AI models into a chaotic, simulated combat aviation environment, yielding results that were as hilarious as they were revealing.

Over on the popular ClaudeAI community on Reddit, a user known as abhi3188 captured the attention of tech enthusiasts by showcasing their latest creative experiment. Previously, this same developer challenged a roster of frontier artificial intelligence models to a boat race, a competition that was ultimately won by Google’s Gemini 4 Argon. Following up on that maritime challenge, the creator decided to up the ante significantly by testing which model possessed the navigational prowess and computational speed required to pilot the iconic aircraft from the film Top Gun: Maverick through its grueling, low-altitude canyon run—a Boeing F/A-18 Super Hornet.

The lineup of participating artificial intelligence models read like a who’s who of the modern tech landscape. The digital competitors included ChatGPT Astra, Google Gemini Carbon, Claude Fable 5.2, and Claude Opus 5.5. To add a definitive baseline for

human-level execution, the artificial intelligence models were measured against the legendary fictional aviator Pete "Maverick" Mitchell himself, represented in the simulation by the parameters of his character’s cinematic flight profile. While Maverick would fly the course precisely as depicted in the hit movie, relying on cinematic intuition and decades of elite naval training, the artificial intelligence systems were given no prior flight instruction or pre-programmed shortcuts. They had to figure out how to pilot a supersonic jet entirely on the fly.

The technical framework of the experiment was designed to test the pure coding and autonomous control capabilities of each model under immense pressure. Each participating model was tasked with writing its own custom autopilot code from scratch. Once the simulation began, the operational parameters were unforgiving. Ten times every single second, the generated autopilot code received a comprehensive update regarding the jet’s complete physical state, encompassing fifty-six raw numerical variables. These parameters included critical telemetry data such as current altitude, airspeed, gravitational G-force load, immediate terrain clearance ahead, the precise location of the next navigational gate, and the altitude ceiling restrictions of the canyon.

To ensure a level playing field, no external flying tips or hand-held adjustments were provided by the human creator. Instead, each artificial intelligence model was granted a preliminary window of ten test flights to analyze its performance, review feedback on what went wrong during previous runs, and fix its underlying autopilot code accordingly. Following this limited iterative debugging phase, every model was given one final, high-stakes shot at completing the full canyon run in real time.

Someone used Claude Code to remake the Top Gun canyon to see which frontier LLMs can outfly Maverick

The resulting performances offered a fascinating glimpse into how different artificial intelligence architectures approach complex spatial challenges and sudden environmental hazards. At first glance, Claude Opus 5.5 appeared to turn in a remarkably poor performance by failing to navigate the course effectively. However, according to the experiment’s creator, abhi3188, the model’s subpar showing was not necessarily a technical failure of computation, but rather a philosophical refusal. Opus 5.5 reportedly balked at the objective, essentially refusing to take control of and pilot a machine of war through a tactical environment.

Meanwhile, the other competitors faced more mechanical downfalls. Both Claude Fable 5.2 and Google Gemini Carbon stumbled right out of the gate, missing the very first navigational checkpoint and resulting in their immediate elimination from the competition. This left ChatGPT Astra as the sole remaining artificial intelligence contender to carry the torch against the cinematic benchmark set by Maverick.

Astra managed to hold its ground—or rather, its airspace—against the human-piloted baseline for an exceptionally impressive duration. The model successfully processed the rapid-fire telemetry stream and maintained control of the Super Hornet through treacherous turns and dramatic altitude shifts. However, the limits of its spatial reasoning eventually caught up with it. Astra misjudged the required distance and clearance while attempting to fly the fighter jet underneath a low-hanging bridge, clipping the structure and driving the multi-million-dollar aircraft straight into the canyon floor in a spectacular crash.

While standard artificial intelligence benchmarks often rely on dry, abstract charts, academic percentages, and standardized multi-choice evaluations provided directly by tech companies during their product release cycles, experiments like this provide a much more entertaining and intuitive metric for measuring progress. They strip away the marketing polish and force models to interpret real-world physics in dynamic environments, highlighting just how far the technology has evolved while simultaneously underscoring the vast chasms that remain unbridled.

Beyond the entertainment value of watching advanced algorithms smash supersonic jets into virtual canyon walls, the experiment also shed light on the economic realities of running complex, high-frequency inference tasks on frontier models. As companies debate the commercial viability and operational overhead of autonomous agents, the financial toll of running these simulations offers an interesting secondary data point.

When the simulation concluded and the API usage tallies were calculated, the cost per flight varied wildly depending on the model utilized. Claude Opus 5.5 recorded an API cost of zero dollars, aligning with its refusal to participate in piloting the combat aircraft. ChatGPT Astra racked up a modest bill of $1.92 for its valiant, albeit ill-fated, attempt to clear the bridge. Google Gemini Carbon cost $2.27 to run through its brief time in the airspace before missing the opening gate. At the higher end of the financial spectrum, Claude Fable 5.2 accumulated an API usage cost of $14.57 during its debugging and execution phases.

As developers continue to push the boundaries of what large language models and autonomous agents can achieve in real-time environments, tests like the Super Hornet canyon race remind observers that artificial intelligence remains a tool of immense potential wrapped in distinct, unpredictable limitations. Until these systems master the art of canyon navigation without turning a fighter jet into a smoking crater, commercial passengers can rest easy knowing that human pilots will remain firmly at the controls.

By Nana Wu