Overview
- Meta launched Muse Code in beta on Wednesday, August 5, 2026, as a terminal tool that it says can plan, write, test and validate complete software engineering tasks across large repositories.
- Muse Code runs persistent background subagents that retain context during a session and records every model call, tool run, approval and edit in a local append‑only event log so interrupted jobs can resume exactly where they stopped.
- The release includes Muse Spark 1.2, a coding‑focused model co‑trained with the agent to improve long‑horizon workflows and complex debugging through extended tool use and planning.
- Meta published pricing that includes a pay‑as‑you‑go tier (reported at $1.25 per million input tokens and $4.25 per million output tokens) plus a deeply discounted contributor tier that requires consent to let Meta use prompts and completions to train future models.
- On Meta’s benchmark charts Muse Spark 1.2 is competitive but generally behind Anthropic’s top model on several coding tests, and Meta says Muse Code is aimed at lower cost and long‑running agentic tasks as it pushes to monetize its AI work and add enterprise options such as reduced data retention on request.