Image: GitHub
Why it mattersA student on the twenty-dollar Claude plan stops hitting the weekly limit as fast. The team on a bigger plan pays a few dollars for DeepSeek instead of Opus tokens.
Opus plans. A model that costs half a cent per task does the typing. A test has to pass before the change reaches you. That is Cheap Agent, an open-source MCP server for Claude Code. The author, who publishes on GitHub as muaaz9329, released it on 28 October. Four days later the repo has 109 stars, seven forks and an MIT licence.
What the project does
Cheap Agent installs as a Model Context Protocol server for Claude Code. The expensive model you already pay for stays in charge. Opus reads the plan, decides what each step should do and asks for the change. Cheap Agent takes that step. A cheap model, in its own copy of your project, writes the code. Then it runs the tests. Only a change that passes the test comes back to Claude.
The author lists four providers for the cheap model: OpenRouter, DeepSeek, OpenCode Go and Command Code. The install is npx cheap-agent. Then in each project, once, /cheapag:allow.
The numbers the author reports
The author ran SWE-bench Verified on Claude Code with Opus 5.5, with and without Cheap Agent. These are the figures stated in the repository README. Nobody else has repeated them yet.
Adding a feature used 154 thousand Opus tokens without Cheap Agent, and 57 thousand with it. That is 63 percent fewer Opus tokens. Both versions passed all their tests, by the author's account.
Seven everyday bug fixes used 1.6 million Opus tokens without Cheap Agent, and 655 thousand with it. That is 59 percent fewer. Both versions solved all seven, by the author's account.
Nine of the hardest bugs used 3.0 million Opus tokens without Cheap Agent, and 1.3 million with it. That is 55 percent fewer. Here Opus alone solved seven of nine. Cheap Agent solved six of nine. A Sonnet helper, in the author's setup, took the leftovers.
Cheap Agent's own cost, across each test, was close to half a cent per task. The cheap model is paid by the user, by their own API key.
Who it is written for
The author's goal, in the README: get the Opus use that costs $200 a month out of the $100 Claude plan. Get $50 to $70 of value out of the $20 plan. Pay a few dollars to DeepSeek for the typing. The author names students in developing countries who want Opus but cannot afford the higher plans.
The repository is TypeScript. It installs on macOS and Linux. The Windows build is a beta. It works in CI but needs Git for Windows. The author lists three Windows-only limits in the README.
What to watch for
The 109 stars at four days are modest. The author's benchmark is the only one so far. The harness is Opus 5.5 against Opus 5.5 plus Cheap Agent. There is no independent run.
The gate is the test. If the project's tests do not catch a wrong-looking patch, the user still gets that patch. The reader then burns the same Opus tokens on review. The author states the SWE-bench numbers plainly. A reader has to decide if their own tests are strong enough to use as the same gate.
Source
muaaz9329/cheap-ag, on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.