Spending Too Much Money on a Coding Agent
10 months ago
- #LLMs
- #AI Coding
- #Productivity
- The author switched from Claude Sonnet to OpenAI o3 for coding tasks, finding o3 more effective despite being slower and more expensive.
- o3 demonstrated improved abilities in using tools, assessing progress, and self-correcting, leading to better results.
- The cost of using o3 was high, averaging $50/day, but the author and co-founder decided to trial a $1000/month budget for o3, finding it worth the expense.
- Large thinking models like o3 and Claude Opus offer advantages such as better tool usage, less tech debt, and more effective research capabilities.
- Tips for maximizing value from coding agents include shifting errors earlier, using boring technology, refining Cursor rules, and improving dev scripts.
- Recent price drops and new plans (e.g., OpenAI's 80% price cut, Cursor's Ultra plan) make large agentic models more affordable.
- New workflows with coding agents include spinning up background agents, scripting refactors, and having multiple agents work simultaneously.
- The future of coding with LLMs shifts from generating unmaintainable code to enabling clear, maintainable improvements with multiple agents.
- Thomas Ptacek's perspective highlights that developers still own curation and judgment, while coding models handle the schlep work.