Meta Muse Code: Beta Launch, Pricing & Agent Features

A one-command install that fans out into subagents

Muse Code is a terminal-based coding agent that Meta Superintelligence Labs put into beta on August 5, 2026. Alexandr Wang, Meta's chief AI officer, described it as installable with one command and able to "take on complete software engineering tasks across a wide variety of use cases, planning changes, writing code, validating the results." It runs alongside Meta's newest model, Muse Spark 1.2, which Wang said was developed and trained alongside the agent itself, improving its coding performance. [2][1]

For large jobs, TechCrunch reported that Muse Code fans out into separate subagents that work in parallel inside isolated git worktrees, so each subagent's changes stay quarantined from the user's working copy until they are ready to merge. In Meta's own testing, the tool built six separate game features at once without the subagents' changes conflicting. [1]

Atlas interpretation: That worktree isolation is the same shape of problem every terminal coding agent in this category has had to solve: letting several agents touch a codebase at once without one agent's half-finished edit corrupting another's. Meta's answer, running each subagent in its own worktree rather than a shared working directory, is an engineering choice rather than a model capability, which is consistent with Wang's own framing of Muse Code as a product built around an existing model rather than a demonstration of a new one. [1]

Priced to undercut, not to differ

Muse Code is available pay-as-you-go at what Wang said is the same API pricing as theMuse Spark 1.1 release: $1.25 per million input tokens and $4.25 per million output tokens. Wang declined to share user numbers for the Muse Spark models but said "adoption has been exciting and strong." He described Muse Code's competitive target directly, telling CNBC, "We think that for a lot of workflows and use cases, this can be an incredibly good option, especially from a cost perspective," naming Anthropic and OpenAI's coding tools as the comparison. [2]

Atlas interpretation: Wang's pitch is explicitly on price rather than on any claimed edge in coding quality, and CNBC frames Muse Code as Meta's first coding agent going up against Anthropic's Claude Code and OpenAI's Codex on those terms. Reusing the Muse Spark 1.1 rate card rather than setting a new one for the agent suggests Meta is treating the token price as already settled and competing on the tool wrapped around it instead. [2]

A revenue product launched the week the market punished Meta's AI spending

CNBC reported that Meta's shares had tumbled the prior week after the company issued a light revenue forecast and disclosed dwindling free cash flow in its second-quarter results, and it frames Muse Code as another route by which Mark Zuckerberg is trying to generate revenue from AI even as the company keeps increasing spending on data centers and related computing infrastructure. Wang, who joined Meta in June 2025 to lead the rebuilt superintelligence effort, oversees Muse Code as part of that foundation model group. [2]

Atlas interpretation: CNBC's own framing ties the two developments together, but the article does not report a specific revenue figure or projection tied to Muse Code, so the connection is about timing and narrative rather than a disclosed financial link. A metered, per-token coding agent is a more direct monetization path than Meta's history of giving model weights away for free, and it arrives from the same lab that closed the weights on Muse Spark itself. [2]