Meta released Muse Code in early beta on 5 August, its first coding agent, running in the terminal on macOS and Linux. It is built on Muse Spark 1.2, which Meta describes as improved on code generation, complex debugging and codebase understanding. The product lands directly opposite Anthropic's Claude Code, OpenAI's Codex and Google's Gemini CLI.
What it does
One command line plans changes, executes them and validates results. The agent delegates to sub-agents, runs persistent background tasks while the developer works elsewhere, and uses isolated git worktrees so parallel tasks do not collide. Meta's demo builds a working webpage from an MP4 file.
The price is the product
Standard pay-as-you-go is $1.25 / $4.25 per million input / output tokens. The Contributor tier is $0.10 / $0.20. Against Anthropic's Sonnet 5 at $3 / $15, the standard tier is already about a third of the price and the Contributor tier is not in the same order of magnitude.
What Contributor costs instead
The discount is conditional: Contributor users must supply feedback that improves the coding agent. Meta has not published what that collection covers, how it is scoped against proprietary repositories, or whether it can be switched off mid-session while keeping the price. For an agent that reads whole codebases, the terms of that exchange are the whole question.
Beta means beta
There is no Windows build, no stated rate limit, no enterprise tier and no published benchmark against Claude Code or Codex. Meta is buying entry to a market whose incumbents have two years of enterprise deployment behind them.
