I think we are living in some of the best time for AI coding agents. Recently, I ditched Claude Code to see if I could make it with OpenCode and I was pleasantly surprised:
- OpenCode has a bunch of free models on Zen
- Ollama has free Gemma 4, Minimax M3
- OpenAI gives you free GPT 5.6 Terra, Luna
Luna is shockingly good, too. I planned a session with Luna High. Then here's the best part. I got Muse Spark to delegate tasks to Gemma 4 according to the plan, so I wouldn't burn too much allowance on GPT free. It worked. A bit of redundant code, though, cause of Gemma 4.
A downside of this approach is that, I had to read code, plan out what and how exactly I would build it. But that's why I'm a software engineer, right? Occasionally, I still code by hand. Even if I mess up by hand, I can ask Google and other AI models on where I get things wrong. It keeps the mind sharp. This whole agent workflow is the result from a rebuild I did for Papersy. I burned through $80 worth of tokens with Claude Opus and Sonnet, while Opus only did task monitoring. I thought Opus fixing bugs from Sonnet subagent coding jobs and running test loop contributed to that large bill. At one point, I had to compact the conversation because it was too long (> 600k tokens).
I thank my job for allowing me to make my own decision.