Meta released Muse Code, a terminal-based coding agent now in beta, together with Muse Spark 1.2, a coding-focused model update available through its paid API. The cheapest pricing tier costs a fraction of the standard rate but grants Meta permission to train on submitted prompts and code, so teams with confidential codebases must choose the more expensive tier. Meta's own benchmark charts place the model second behind Anthropic's Opus 5, and nothing is offered as open weights.
What changed
Meta's developer offering centred on free downloadable Llama weights, with no coding agent of its own.
What it unlocks
Running a terminal-based coding agent from Meta that keeps helper agents alive through a session, works in parallel isolated copies of a repository, and resumes long jobs after a crash from a local log.
- contributor tier $0.10/$0.20 per 1M tokens
- standard tier $1.25/$4.25 per 1M tokens
- 60 vs 3,000 requests per minute
- 82.9% Terminal-Bench 2.1, 2nd to 86.7%
What you need to act on it
- Meta account with billing details on file
- macOS or Linux
- standard paid tier to keep code out of Meta's training data
Sources