What they did
The poster kept running out of their Opus allowance in Claude Code and saw that most of it went on mechanical work: reading files, writing boilerplate and rerunning tests. They built Cheap Agent, an MCP server and Claude Code plugin. Opus plans and writes a short task brief; DeepSeek V4.1 Flash, called through OpenRouter, does the edit in a separate git worktree. The cheap model has to add a test that fails before its change and passes after, and the plugin runs that test itself rather than trusting the model. Only passing work is applied, left uncommitted for a normal diff review. If the cheap model stalls, a Sonnet helper finishes the step. The poster had Opus design the tool and then used Cheap Agent to build the rest of it.
How it went
On SWE-bench Verified with Opus 5.5 and the same prompts, a feature task dropped from about 154k to 57k Opus tokens. Seven everyday bug fixes used 59% fewer Opus tokens and both setups solved all seven. On nine of the hardest bugs it saved 55% but solved six against Opus alone's seven. The Windows port was 30 delegated jobs costing $0.29 in DeepSeek calls. It gave no gain on tiny jobs, and an early demo saved only 20% because the repo had no test command, so Opus ran the tests itself.
Worth knowing
The savings depend on the repo having a working test command. The poster found that letting Opus decide what counts as simple failed, because it kept doing small edits itself; sending every code change to the cheap model worked better. Code goes to whichever provider you configure. Several commenters pointed out that subagents on a smaller Claude model do much the same, and one reported a long migration that a cheaper model kept failing until Opus took over.