This video presents a head-to-head comparison of Google's Gemini 2.5 Pro (with the newly released Deepthink feature) and OpenAI's Chat GPT5 Pro in a complex code generation task. The objective was to prompt both AI models to create a web-based 3D skateboarding physics simulator. Despite being cutting-edge models with high subscription costs, both failed to produce fully functional game code. Chat GPT5 Pro's initial output was slightly more functional (supporting kickflips but lacking movement), while Gemini Deepthink's was less so. Attempts to have the models self-correct also proved largely unsuccessful, with GPT5 Pro showing no measurable improvement and Gemini Deepthink encountering persistent errors. The video concludes that while these models are advancing, they currently struggle with intricate, multi-faceted coding challenges without iterative human intervention or an agent-based framework.
Comparing AI Models for Game Development
0:03 Introduction & Purpose: The video's main goal is to test Google's Gemini 2.5 Pro (with the new Deepthink feature) against OpenAI's Chat GPT5 Pro, both recently updated or released, in a complex code generation task.
0:51 High-Cost, Limited Access: Both models are locked behind expensive subscriptions ($250/month for Gemini Deepthink with limited usage, $200/month for Chat GPT5 Pro).
1:40 The Prompt: 3D Skateboarding Simulator: The models were tasked with generating an "advanced web-based 3D skateboarding physics simulator," including design, physics implementation, and game mechanics for a rectangular skate park with a third-person view and trick abilities.
2:59 Simultaneous Execution: Both models were given the exact same prompt at nearly the same time to ensure a fair comparison.
4:27 Deepthink's Speed Advantage: Gemini 2.5 Pro Deepthink finished generating its response significantly faster than Chat GPT5 Pro, which took 12 minutes and 25 seconds for both responses to be ready.
5:01 Blind Testing Methodology: The presenter implemented a blind testing method by saving the generated HTML files with random names, ensuring an unbiased initial evaluation.
6:02 GPT5 Pro's Initial Output (First Test): Identified by a "depressing color palette," the game loaded, displayed shadows and a staircase, and the "kickflip" function worked (showing "in air" state), but forward/backward movement was non-functional.
7:29 Gemini Deepthink's Initial Output (First Test): Although visually similar, this version was "significantly worse," lacking kickflip and jump functionality, and overall non-functional.
8:57 GPT5 Pro Performed Better Initially: Contrary to the presenter's previous experiences, Chat GPT5 Pro did a "significantly better job" than Gemini Deepthink in the first attempt.
9:59 Attempts to Fix the Code: The presenter tried to prompt both models to fix their generated code. Gemini Deepthink required starting a new chat for another Deepthink usage.
11:29 Re-evaluation of Fixed Code (GPT5 Pro): Chat GPT5 Pro apologized and identified a "ground raycast bug." However, its "fixed" code still failed to enable proper steering or movement, though some side-to-side balancing was observed.
11:03 Gemini Deepthink's Failure to Fix: Gemini Deepthink encountered a "string of errors" (e.g., "something went wrong," "cannot complete this request") and was unable to generate a corrected script.
12:54 Conclusion: Both Models Struggle with Complex Code: The testing concluded with dissatisfying results, as neither model could produce a fully functional game nor effectively self-correct.
14:02 Future Implications for AI Coding: The presenter suggests that while AI models are improving, complex coding tasks may still require an agent-based framework that allows for iterative refinement and self-correction, rather than single-shot generation.