This video presents a hands-on first look at OpenAI's recently released GPT-5 Pro model, specifically evaluating its advanced coding capabilities through three distinct generation tasks. The presenter tests the model's ability to create a browser-based operating system, a web-based 3D low-poly racing game, and a Python-based first-person shooter.
The GPT-5 Pro model demonstrates exceptional performance in generating a highly functional and intricately designed web operating system, featuring movable/resizable windows, a robust terminal with extensive commands, a working calculator, a file explorer with file creation/deletion, and customizable themes—surpassing any previous results observed by the presenter. However, its performance on the 3D racing game was notably poor, plagued by camera and movement issues, falling short of competitors like Gemini 2.5 Pro Deepthink. Conversely, the Python FPS game, despite some minor glitches, was surprisingly good and enjoyable, featuring procedural map generation and basic gameplay mechanics. The video concludes that GPT-5 Pro exhibits significant leaps in code generation complexity and functionality, particularly in UI/OS and game development, while acknowledging some inconsistencies.
Hands-on Test of GPT-5 Pro's Coding Capabilities
0:05 GPT-5 Pro Release: OpenAI's GPT-5 Pro model has been released and is rolling out to users, prompting a hands-on testing session.
0:21 Testing Methodology: The presenter utilizes the Pro variant of GPT-5 to test prompts typically reserved for state-of-the-art models, noting OpenAI's inclusion of game demos in the announcement, suggesting a focus on software engineering capabilities.
Prompt: Generate a browser-based operating system with elements like a functional terminal, calculator, text editor, clock, and file browser.
Result (03:09): GPT-5 Pro produced an "Oh, wow" response from the presenter, showcasing significant improvements in UI and functionality.
Key Features (03:31): Movable windows, functional calculator, a terminal with an extensive list of commands (e.g., pwd, ls, cd, mkdir), functional file explorer with file creation and deletion, customizable desktop backgrounds (including custom images), and theme changes.
Key Takeaway (03:52): This was by far the "best result ever seen" for this test, demonstrating a significant generational leap in intricacy and functionality compared to previous models.
8:22 Web-Based 3D Racing Game Test (Disappointing):
Prompt: Create a low-poly 3D racing game with a cockpit view, simple track, opponent, and lap counter.
Result (10:06): The model provided a download link for an HTML file instead of in-line code.
Issues (10:42): The initial camera Z-height was too low, making the game unplayable without manual code adjustments. After adjustment, the car only moved backward, the opponent car behaved erratically, and the "road is a disaster," with an "interesting, somewhat disturbing color palette."
Key Takeaway (11:00): This test yielded a "pretty poor" result, performing worse than Gemini 2.5 Pro Deepthink for this specific task, despite the strong performance in the OS test.
12:47 Python-Based First-Person Shooter Test (Very Good):
Prompt: Create a first-person shooter using Python with full creative freedom, aiming to "wow" the presenter.
Result (13:45): The game featured procedural map generation (N for new map), basic movement (W A S D), sprint (Shift), mouse look, and firing (Left Click).
Functionality (14:10): Ammunition count, weapon switching (1 and 2), health packs, and enemies that attack (though sometimes hard to hit).
Challenges (13:59): Initial spawn issues (stuck in walls), occasional lagging, and some glitches, but overall "awesome" and "impressive" given the procedural generation.
Key Takeaway (16:34): Despite minor glitches, this result was "really good" and enjoyable, making up for the racing game's failure.
17:17 Overall Conclusion: GPT-5 Pro works "pretty decently," with the web-based OS being an "absolutely beautiful" and "most intricate" result, representing a true generational leap. The FPS game was "very good" for a one-shot test, but the 3D racing game was a noticeable weakness.
18:51 Future Content: The presenter encourages questions, subscriptions, and anticipates more interesting findings as time progresses.