Meta launches Muse Code terminal agent, undercuts Claude Codex on price

Meta has released Muse Code, its first coding agent built to compete directly with Anthropic's Claude Codex and OpenAI's equivalent tools. The terminal agent, available in beta, can plan changes, write code, and validate results across a wide range of software engineering tasks. It works alongside Muse Spark 1.2, Meta's latest coding-specialized model that was co-developed with Muse Code for better performance together.

Pricing is the primary differentiator. Pay-as-you-go access costs $1.25 per million input tokens and $4.25 per million output tokens—undercutting comparable frontier models. A contributor tier available for developers willing to opt-in to help improve the model costs more than 10 times cheaper than the standard tier, following industry-standard model-improvement practices.

The agent was announced by Alexandr Wang, Meta's Chief AI Officer and head of Meta Superintelligence Labs, who oversees the company's foundation model strategy. Wang emphasized that developers can swap models and harnesses if they want, but using the Meta-built components together yields best performance. Meta is also offering zero-data-retention options on request—a feature targeted at enterprise customers concerned about data retention for compliance. For architects evaluating coding assistants, Meta's pricing is now a table-stake factor alongside performance, and the co-development story (agent + model trained together) signals a shift toward integrated toolchains rather than modular components.