cheapAg : My Opus was hitting limits on boilerplate, so I made a $0.005 model do the typing
retained text
I have been working with Claude for a really long time (Aug 2024 , the copy pasting era lol) and alwaya have loved Opus and it's ability to think about problem and then provide options , overall the thinking and planning ability of Opus/Fable is something that no other model can tops But From the start of Agentic Era I kept hitting my Opus limit, and most of it was going on typing: reading files, writing boilerplate, re-running tests , something that very cheap model can do very well as well but never got time to make this Framework/workflow properly So I built **Cheap Agent**, an MCP server and plugin for Claude Code. **What it does** * Opus plans and writes a short task. * A cheap model (DeepSeek V4.1 Flash via OpenRouter, about half a cent a task) writes the code in a git worktree. * It has to write a test that fails before the fix and passes after. Cheap Agent runs that test itself. * Only passing work lands, uncommitted, so you review the diff as usual. * If the cheap model gets stuck, a Sonnet helper finishes the step instead of Opus. **Numbers** (SWE-bench Verified, Opus 5.5, same prompts with and without it): * Adding a feature: 154k → 57k Opus tokens, 63% less * 7 everyday bug fixes: 59% less, 7/7 solved by both * 9 of the hardest bugs: 55% less, but 6/9 solved vs 7/9 for Opus alone It doesn’t help on tiny jobs, because Opus is already quick there. **Why I made it** I’ve met a lot of students who want Opus and can’t afford $100 a month. The aim is to stretch the plan you already pay for. **How It Was build :** in Claude Code. Opus designed it, then I used Cheap Agent to build Cheap Agent (sounds funny lol) . The Windows port was 30 delegated jobs for $0.29 of DeepSeek. cheap models are fine at typing as long as the task is tight, but never trust one that says "tests pass", so Cheap Agent runs the tests itself. And the savings depend on your repo having a test command: my first demo take only saved 20% because I forgot a package.json, so Opus ended up running the tests itself. **Free Forever:** it’s open source (MIT) and free. The cheap model runs on your own API key, and a few dollars lasts hundreds of tasks. Your code goes to the provider you pick, and only after you allow it per project. `npx cheap-agent` · [github.com/muaaz9329/cheap-ag](http://github.com/muaaz9329/cheap-ag) would really love your guys feedback and how can I improve this , better yet could have contributions from Awesome community here on what people might need Hope everyone have great weekend ! **Edit, answering the two common questions:** **vs Opus + Sonnet:** Sonnet still uses your Claude limit. Here the work runs on a model outside your plan for \~$0.005 a task, and Sonnet is only the fallback. **vs an Opus → DeepSeek loop:** the cheap model also does the *reading*. Opus asks it questions and plans from short answers instead of opening files. That reading is where most Opus tokens go.
document fetched 2026-10-31 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1xjqoi0/cheapag_my_opus_was_hitting_limits_on_boilerplate