Meta entered the AI coding-agent race on August 5, with Mark Zuckerberg announcing Muse Code in beta on X that evening, followed the next morning by a fuller rollout from the company's AI team. Muse Code is a terminal-based agent built to handle complete software engineering tasks across large repositories — planning a change, writing the code, running it, and validating the result — rather than acting as an autocomplete or chat sidebar bolted onto an editor. It's Meta's first standalone product aimed squarely at the agentic coding tools that have driven much of Anthropic's and OpenAI's recent growth.

What Muse Code does

The agent is powered by Muse Spark 1.2, a new version of Meta's model tuned specifically for "code generation, complex debugging, codebase understanding, and end-to-end developer workflows." Its headline feature for large codebases is a 1-million-token context window, which lets Muse Code load an entire project into a single session instead of chunking it into smaller pieces the way many coding agents have to. For bigger jobs, Muse Code can spin up its own persistent, asynchronous sub-agents and delegate parts of the work to them so multiple pieces of a task run in parallel. An event-history log records every action the agent takes during a run, so a long, multi-stage task can resume from where it left off after a failure instead of starting over. The beta is available for macOS and Linux.

Pricing and positioning

Meta is offering Muse Code through a two-tier, pay-as-you-go pricing structure. A Standard Tier charges $1.25 per million input tokens and $4.25 per million output tokens (cached input at $0.15 per million), with Meta committing that prompts and completions on this tier aren't used to train its models — pricing in line with what Meta charged for API access to the prior Muse Spark 1.1 model. A cheaper Contributor Tier drops those rates to $0.10 per million input tokens and $0.20 per million output tokens, in exchange for Meta being able to use that usage data for model training.

Why it matters

Coding agents have become one of the most competitive fronts in AI, with Anthropic's Claude Code and OpenAI's Codex both credited with driving significant subscription and API revenue this year. Muse Code is Meta's answer, and the design choices — a terminal-first workflow, parallel sub-agents, and a very large context window — signal that Meta is betting on handling large, real-world codebases as its point of differentiation rather than competing purely on raw benchmark scores. Whether Muse Spark 1.2's coding ability actually matches Claude Code or Codex in practice will only become clear as independent developers put the beta through its paces on real projects.