TLDR;
This video provides a first look at OpenAI's new coding model, GPT5 Codeex, a refined version of GPT5 optimized for software engineering. The model is designed for agentic programming, offering faster performance, improved reliability, and better real-time collaboration. The key takeaways include:
- GPT5 Codeex dynamically adjusts its "thinking time" based on task complexity, leading to more efficient token usage.
- It demonstrates improved performance in code refactoring and generates more meaningful code review comments.
- The model provides more succinct and direct responses, streamlining the development process.
Intro [0:00]
The presenter introduces OpenAI's new coding model, GPT5 Codeex, highlighting its potential as a condensed and refined version of GPT5 specifically designed for software engineering tasks. He expresses his initial positive impressions and outlines the video's plan to explore the model's announcement and practical usage.
Announcement [0:28]
OpenAI has released GPT5 Codeex, a version of GPT5 optimized for agentic software engineering. It is faster, more reliable, and better at real-time collaboration. GPT5 Codeex achieves better scores on the Swebench benchmark compared to GPT5 High and excels in code refactoring tasks. A key feature is its ability to dynamically adjust thinking time based on task complexity, providing quicker responses for simple requests and more thorough processing for complex tasks. The model has been observed working independently for over 7 hours on large tasks, iterating on implementations and fixing test failures. Token usage is more efficient, with simpler tasks requiring significantly fewer tokens and complex tasks utilizing more. GPT5 Codeex is also tuned for code reviews, reducing incorrect comments and focusing on high-impact, meaningful feedback.
Using in CodexCLI [4:13]
The presenter demonstrates GPT5 Codeex in the terminal using CodexCLI, showcasing a new animation announcing the model. Selecting the GPT5 Codeex model allows users to interact with it within their applications. The model provides faster responses and communicates information differently compared to previous versions of GPT5, offering succinct and conversational readouts with embedded links. In contrast to GPT5 medium, which returns extensive information in a blog post format, GPT5 Codeex delivers more direct and engineering-focused responses, highlighting essential aspects.
Putting it to work [6:34]
The presenter shares his "numbers application" to demonstrate GPT5 Codeex's practical use, addressing a missing Firestore index issue. Firestore, a cloud database solution from Google, requires proper indexing for efficient data filtering and sorting on the cloud. The presenter aims to use the Codeex model to resolve this problem, preventing client-side data handling.
Use shell to fix a problem [7:25]
The presenter uses the cursor editor, and demonstrates two ways to access Codeex: through the terminal panel and via the AI pane with the Codeex extension. He inputs the error message "Missing fire store index limited video query" into Codeex, tasking it with fixing the problem without additional context. Codeex analyzes the issue, referencing relevant parts of the video system and applying more extensive processing. The model generates a composite index definition and provides the next step: running Firebase Fire Store indexes against the Firebase project to install the index. Codeex then opens a shell and executes the command, successfully deploying the index after some iteration.
Visual application test [11:38]
The presenter tests GPT5 Codeex with an older visual project to see if it can update it effectively. He uses a sample application and aims to enhance the filter button to light up when filters are applied, indicating that the displayed content is being filtered.
Cursor Codex panel [12:39]
The presenter uses the Codeex side panel in cursor, selecting the GPT5 Codeex medium model. He provides a screenshot and instructs the model to change the filter button's color to the main accent color when filters other than the default are applied. Codeex successfully implements this change, lighting up the button when filters deviate from the defaults.
Conclusion [13:56]
The presenter summarizes his initial experience with GPT5 Codeex, highlighting its ability to provide a smaller, more efficient model for common tasks while still offering deep thinking capabilities when needed. He appreciates the model's potential to streamline development and expresses interest in user feedback on their experiences with GPT5 Codeex.