GPT-5 vs. Sonnet: Complex Agentic Coding
16 days ago
- #AI
- #coding
- #comparison
- OpenAI released GPT-5, promoting it as their best model for agentic coding.
- GPT-5 was tested against Claude 4 Sonnet for porting a TypeScript tool (Ruler) to Rust.
- GPT-5 demonstrated high intelligence, agency, and precise instruction-following but produced messy code.
- Claude 4 Sonnet was faster and produced elegant, maintainable code but was less disciplined and required more iterations.
- GPT-5 got stuck twice but recovered, while Claude never got stuck but had unreliable status reports.
- GitHub Copilot Chat requires manual approval for terminal commands, limiting autonomy.
- Both models performed well, with GPT-5 excelling in intelligence and Claude in code elegance.