Meta introduces Muse Code, its terminal-based coding agent

Meta has released an early beta version of Muse Code, a terminal-based coding agent targeting Anthropic's Claude Code and OpenAI's Codex. The new tool runs on Muse Spark 1.2, an updated edition of the company's AI system that reportedly delivers "improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows."

Muse Code manages complete software engineering jobs through plain text instructions. The system writes code, maps out modifications, checks outcomes and supports construction of functional programs. It additionally coordinates several sub-agents while assigning duties to speed up delivery.

Meta highlighted demonstration projects that include an interactive photon sphere simulation plus a Plants vs. Zombies-style clone. A video walkthrough depicts the agent creating a full webpage from an MP4 video clip.

Affordability forms a central distinction against rival offerings. By default Muse Code follows the same usage-based rates as Muse Spark, set at $1.25 for every million tokens received and $4.25 for every million tokens produced. According to The Wall Street Journal, a contributor tier charging $0.10 for every million tokens received and $0.20 for every million tokens produced will be available to those who consent to supplying feedback that refines the agent. In a CNBC interview, Meta Chief AI Officer Alexander Wang described the tier as one that gets participants "in at a significantly lower cost."

These figures sit well below Anthropic's Sonnet 5 model, which generally runs $3 for every million tokens received and $15 for every million tokens produced. Should the tool deliver strong results, enterprises may migrate toward Meta. Comparable economics have already prompted numerous firms to adopt Chinese models such as DeepSeek instead of increasingly expensive American alternatives.