AI
Meta's Muse Code Is Drastically Cheaper Than Claude—but Delivers Less Polished Work
Meta's new AI coding agent undercuts Claude by 150x on default pricing, but independent testing reveals a quality tradeoff that may mean more review work for engineers.
The New Stack
Meta launched Muse Code on August 5, a new AI-powered coding agent built on its Muse Spark 1.2 model. The tool promises to handle "complete software engineering tasks across large repositories," with work distributed across parallel sub-agents. Meta's AI leadership emphasized a single competitive advantage: aggressive pricing.
The default Contributor tier costs $0.10 per million input tokens and $0.20 per output million—more than 10x cheaper than Anthropic's Claude Code on the Fable 5 model ($10 input, $50 output). There's a catch: code submitted on the Contributor tier becomes Meta training data by default.
To test whether the dramatic cost savings come at a quality penalty, an independent reviewer ran identical coding tasks against both systems and measured output quality, token efficiency, and actual cost.
Methodical Testing Across Three Real-World Scenarios#
The experimenter used dayjs, a 2KB JavaScript date library with a 773-test suite, to avoid synthetic benchmarks. Three separate tests were designed to measure different aspects of code agent capability:
- Bug identification and repair — fixing three deliberately planted subtle errors without touching test files
- Codebase refactoring — restructuring a monolithic 467-line file while maintaining all tests and linting standards
- Timezone reasoning — diagnosing why tests failed in certain timezones and determining the proper fix without sacrificing test rigor
Both agents operated under identical constraints and fresh sessions for each task.
Bug Hunt: Tied on Surface, Diverged on Completeness#
Both Muse and Claude identified and fixed all three injected bugs correctly, explaining the issues accurately. Where they differed was on handling pre-existing environmental failures unrelated to the planted bugs. Muse correctly identified them as non-issues and left them. Claude went further, engineering a solution that made the entire test suite pass cleanly.
Cost and speed: Muse completed the task in 2 minutes 58 seconds using 2.25 million tokens for 2 cents (Contributor pricing). Claude took 5 minutes 31 seconds and 1.46 million tokens for $3.02.
