At my day job, the company covers Claude Code, Cursor, and Codex. Combined it's around $1,800/month in AI credits. I never really thought about the cost because I wasn't paying for it.

Then I started building my own stuff on weekends and realised I absolutely cannot spend that kind of money as a solo dev.

So over the last few weeks, I've been experimenting with cheaper alternatives. Here's what actually works:

OpenCode, an open source CLI agent, is basically Claude Code, but you bring your own models. $5/month plan gives you access to a bunch of models with solid usage limits.

Models I use inside it:
- Kimi 2.6 for frontend tasks and general coding, surprisingly good.
- DeepSeek V4 for backend and anything that needs real reasoning, genuinely competitive with Claude Sonnet
- Qwen 3.7 as a backup when I need something different

Codex - $20/month for GPT-5.5 access, I use this specifically for brainstorming and architecture thinking, not for writing code

Total monthly spend: $25. I built an entire web app with 300+ users on this stack and didn't even use 10% of my monthly credit limit.

The thing that surprised me most is that once you figure out which model is good at which task, you stop missing Claude Code. The gap between open source and closed source has genuinely shrunk.

Happy to answer any questions about the setup. Also made a short video walking through the exact workflow if anyone wants the visual breakdown.