OpenAI launches "genius" o4 model with a programming CLI tool...

FireshipAbout 3 min readMay 27, 2025Watch original
THE SUMMARYAI-generated

Key Concepts:

  • OpenAI's new reasoning models (03 and 04 Mini)
  • Codeex (OpenAI's open-source CLI tool)
  • AI-powered code generation and analysis
  • Developer tooling landscape (VS Code, Cursor, Windsurf, Firebase Studio, Claude Code, Copilot, Devon, Augment, Bolt)
  • Model Context Protocol Servers
  • Gemini 2.5
  • Firebase Studio (formerly Project IDX)
  • Svelte 5 with runes
  • MX (video infrastructure API)
  • AI slop

1. Introduction and OpenAI's New Models:

  • OpenAI released new reasoning models, 03 and 04 Mini, claimed to be at or above genius level.
  • The release follows shortly after GPT4.1, 40 image gen, and GPT4.5.
  • The focus is on 04, not 40.

2. Codeex: OpenAI's CLI Tool:

  • OpenAI released Codeex, an open-source CLI tool similar to Claude Code.
  • Codeex can write, execute, and analyze code directly from the terminal or IDE.
  • Installed via npm and requires an OpenAI API key.

3. The AI Developer Tooling Arms Race:

  • A massive competition is underway in Silicon Valley to attract software engineers with AI-powered tools.
  • The "code shovel business" is booming.
  • Rumor: OpenAI is in talks to buy Windsurf for $3 billion.
  • Cursor, a VS Code fork, generates $100 million in annual revenue.

4. Microsoft's Copilot and Gemini 2.5:

  • Microsoft's goal is to "embrace, extend, and extinguish" coders (in a positive way).
  • Copilot received a major upgrade called "agent mode," seen as a competitor to Cursor and Windsurf.
  • Copilot can create files, run commands, and integrates model context protocol servers.
  • Gemini 2.5 is considered by many as the best programming model.

5. Firebase Studio:

  • Google released Firebase Studio (formerly Project IDX), a browser-based VS Code fork.
  • Firebase Studio generates code with Gemini 2.5 and can host and deploy code automatically.

6. Testing 04 Mini and Codeex:

  • The presenter attempts to use Codeex to build a YouTube clone.
  • The model requests clarification on the requirements.
  • The process is slow, and the result is a bunch of empty directories.
  • The issue is attributed to using Windows for development.
  • The model fails to write Svelte 5 code with runes.

7. Comparison with Claude Code and Firebase Studio:

  • Claude Code also takes a long time but can run commands on Windows.
  • Claude Code also fails to write Svelte 5 rune code.
  • Firebase Studio is significantly faster.
  • Firebase Studio ignores the request to generate Svelte 5 code and writes in Next.js.
  • AI integration is easier within an IDE like Firebase Studio.

8. Final Verdict and Call to Action:

  • All the AI tools "kind of suck in their own special ways."
  • Don't believe the hype or the doomsayers.
  • Embrace the current state of AI development and "ship as much AI slop as you possibly can."

9. MX: Video Infrastructure API (Sponsor):

  • MX provides API-first video infrastructure for hosting, encoding, real-time analytics, thumbnail generation, and live streaming.
  • MX scales from startups to large companies like Substack, Patreon, and HubSpot.
  • Developers can try MX for free at mx.com/fireship.

10. Conclusion:

  • The developer tooling landscape is chaotic but exciting.
  • AI-powered tools are still imperfect but offer potential.
  • Developers should experiment and ship code.

AI summaries can miss context or contain errors. Check important details against the original video.

Go a little deeper.

Have a question about this video? Load its transcript to open the video chat.