OpenAI Just Dropped A New Coding Model for Developers

Key Concepts:

• GPT5 Codeex: A specialized, optimized version of GPT5 specifically trained for coding tasks.
• Long Horizon Coding Tasks: Complex coding tasks that can run independently for extended periods (e.g., over 7 hours).
• Dynamic Reasoning Effort: The model's ability to dynamically adjust the amount of "thinking time" or computational effort based on the complexity of the task.
• Code Reviews: The process of reviewing code for flaws and correctness, a key focus area for GPT5 Codeex.
• Sweepbench Verified: A benchmark used to evaluate the performance of coding models.
• Token Generation: The process of generating tokens, which are the basic units of text used by language models.
• Codex CLI: A command-line interface for accessing and using GPT5 Codeex.
• Local to Web Handoff: The ability to seamlessly transition tasks between a local environment (e.g., CLI) and a web-based environment.

1. Introduction of GPT5 Codeex

• OpenAI has released GPT5 Codeex, a specialized version of GPT5 trained for coding, accessible within the Codeex environment (terminal, IDE, web).
• It excels at code reviews and handles long horizon coding tasks (over 7 hours independently).
• The speaker notes the confusing naming convention.

2. Availability and Access

• GPT5 Codeex is available in a cloud version within ChatGPT, including a code review version.
• Access is provided through CLI or IDE extensions (VS Code and Cursor).
• It's included in ChatGPT Plus, Pro, Business, Education, and Enterprise plans with rate limits.

3. Benchmarks and Performance

• Sweepbench Verified: GPT5 Codeex shows a slight improvement over GPT5 on high settings.
  • The speaker points out that the benchmark is now tested on all 500 questions, unlike previous tests on only 477 questions.
  • GPT5's performance decreases when tested on all 500 questions, and GPT5 Codeex doesn't surpass the original GPT5 score on the limited question set.
• Code Refactoring: GPT5 Codeex is significantly better (50% improvement) than GPT5 on high settings for code refactoring tasks.
• Token Usage:
  • For the bottom 10% of user turns (low complexity), GPT5 Codeex uses 94% fewer tokens than GPT5.
  • For the top 10% (high complexity), GPT5 Codeex spends twice as long and generates twice the number of tokens, indicating more "thinking time."
  • This demonstrates the model's ability to dynamically select reasoning effort based on task complexity.

4. Dynamic Reasoning and Task Handling

• GPT5 Codeex dynamically adjusts its "thinking time" based on task complexity.
• It combines interactive sessions with developers and independent task execution for longer tasks.
• Users can start tasks locally (CLI) and offload them to the cloud version.

5. Code Review Capabilities

• GPT5 Codeex is specifically trained for code reviews and finding critical flaws.
• It navigates codebases, reasons through dependencies, and runs code/tests for validation.
• Internal benchmarks show a substantial reduction in incorrect comments (from 13% to 4%) and an increase in high-impact comments (by 12-13%).
• The average number of comments per pull request (PR) has also decreased.

6. Integration and User Experience

• Updates are available in Codeex CLI and IDEs (VS Code, Cursor).
• The speaker suggests that Codeex CLI provides the "purest" experience.
• Early access users have reported positive experiences.

7. Key Insights from Dan (Every CEO/Co-founder)

• Highlights two important aspects:
  • Dynamic Thinking Time: The model can decide how much effort to spend based on task complexity.
  • Local to Web Handoff: Seamless transition between local and web environments.

8. Security and Safety

• Codeex runs in a sandbox environment with network access disabled for safety.
• This prevents malicious code execution or destructive actions.

9. Pricing and Availability

• Included in ChatGPT Plus, Pro, Business, Education, and Enterprise plans.
• Generous rate limits are provided, with Pro plans supporting a full work week.
• GPT5 Codeex is not yet available through the API but is planned for future release.

10. Conclusion

• GPT5 Codeex represents a significant advancement in AI-powered coding assistance, particularly in code reviews and handling complex, long-horizon tasks. Its ability to dynamically adjust reasoning effort and seamlessly transition between local and web environments offers a powerful and flexible user experience. The speaker plans to test GPT5 Codeex in Codeex CLI and IDEs and report the findings in a future video.