Does GPT-5 live up to the hype? [Real Coding Test]

Key Concepts

GPT5, Cursor, coding model, Product Hunt, landing page update, full-stack app development, farming simulation game, Volo app, JSON blob, game state, graphics implementation, development phases, Claude models, Cursor CLI.

Updating a Landing Page with GPT5

The video begins with the presenter testing OpenAI's newly released GPT5, integrated into the Cursor IDE, by making changes to a real-world application. The initial task involves updating a landing page for a product launch on Product Hunt.

1. Initial Prompt: The presenter provides GPT5 with a prompt detailing six specific areas of the landing page to be modified.
2. GPT5's Execution: GPT5 analyzes the project, reads relevant files, and modifies seven files.
3. Initial Results and Issues: The initial changes are not satisfactory. The presenter notes that GPT5 took a long time to make relatively few changes and failed to fully grasp the intended design for a specific section. The UI was broken in some parts.
4. Follow-up Instructions: The presenter provides clarifying instructions, including a specific Product Hunt embed link, style adjustments, and the addition of a YouTube video section.
5. Further Modifications: GPT5 makes additional changes based on the follow-up instructions.
6. More Issues and Backtracking: The presenter identifies new issues, such as unwanted color changes and incorrect border styles. He requests GPT5 to backtrack on some changes and add further details.
7. Iterative Refinement: GPT5 successfully backtracks on some changes and adjusts colors, but other elements require further tweaking.
8. Observation on Model Behavior: The presenter observes that GPT5 tends to do more than explicitly asked, a characteristic previously disliked in Sonnet 37. He notes that Sonnet 4 was better at following instructions precisely.
9. Final Adjustments: After about 25 minutes and multiple prompts, the presenter achieves the desired landing page design. However, he notes an inconsistency where a call-to-action (CTA) was updated in one location but not another, a mistake he believes Sonnet would have avoided.
10. Specific Issue: The presenter mentions an error at the end of the process where GPT5 struggled to make a countdown timer start immediately upon page refresh, a problem Sonnet solved quickly.

Building a Full-Stack Farming Simulation Game with GPT5

The second part of the video focuses on building a simple full-stack farming simulation game from scratch using GPT5.

1. Project Initialization: The presenter uses npx create volo app dot to initialize a full-stack application using the Volo app starter kit. The Volo app is described as a free, open-source starter kit that handles initial setup tasks like database and authentication.
2. Project Brief Creation: The presenter uses the "create brief" feature in Cursor to define the project's scope and goals. The game is described as a browser-based farming simulation where players can move around, farm, feed animals, and decorate their house.
3. Planning Phase: The presenter uses the "plan feature" command to outline the necessary steps for building the initial version of the game. He specifies that the game state should be saved in a database to make it a full-stack application.
4. Initial Plan Review: GPT5 generates an initial plan, which the presenter finds somewhat dense and lacking proper markdown formatting. He notes a suggestion to use a single JSON blob for the MVP to store all game state, which he considers a questionable but potentially efficient decision. The initial plan also lacks details on graphics implementation.
5. Adding Graphics Details: The presenter asks GPT5 to add details on graphics implementation, which it does.
6. Implementation Phases: The presenter initiates the implementation phase, starting with phase one. He notes that GPT5 is slow, taking 5-10 minutes to write approximately 100 lines of code.
7. Parallel Development: To expedite the process, the presenter opens multiple tabs and instructs GPT5 to work on different phases of the project in parallel (phases two through seven).
8. Final Result: After about an hour and a half, GPT5 completes six phases of development. The presenter reveals the game, which he finds unimpressive. He expresses a feeling that Claude models could have achieved better results.
9. Issues Encountered: The presenter notes that GPT5 encountered errors during development, requiring him to repeatedly paste code back in. He also criticizes GPT5 for overengineering certain aspects and writing legacy migrations despite the project being new.

Conclusion

The presenter concludes that GPT5 was a "bit of a dud" based on his first 24 hours of experience. He found it slow, prone to errors, and prone to overengineering. He states that he will continue using Anthropic's Claude models for now. He also expresses interest in re-evaluating Claude Code and exploring new features in the Cursor CLI. He encourages viewers to share their experiences with GPT5 in the comments.