Meta Ships Muse Code, Its First Coding Agent, on a New Muse Spark 1.2 Model

Meta released an early beta of Muse Code, a command-line coding agent, together with Muse Spark 1.2, a coding-focused model the company says it co-trained with the agent so that model and harness fit each other. Meta describes gains over Muse Spark 1.1 in code generation, complex debugging, and codebase comprehension, which it attributes to more coding training compute and a wider set of training environments.

The agent's design choices are the more interesting part. Muse Code installs from the terminal with one command, runs several subagents in parallel with implementation and review happening at the same time, and keeps a replayable event log of every subagent, tool call, intervention, and cancellation. VentureBeat reported persistent asynchronous background agents as the headline capability, and coverage put the model's context window at one million tokens. That is an argument for auditability, directed at engineering managers who will not let an unlogged agent touch a production repository.

Pricing is where Meta is doing something other labs have avoided stating explicitly. The standard muse-spark-1.2 rate on the Meta Model API is $0.15 per million cached input tokens, $1.25 per million input tokens, and $4.25 per million output tokens. Alongside it sits a "contributor" tier at a materially lower price in exchange for Meta using the user's prompts and completions as training data. That converts a privacy preference into an explicit line item.

No independent benchmark reproduction has appeared yet. Meta's claims of improvement over version 1.1 are internal, and the comparisons that matter, against Anthropic's Claude Code and OpenAI's coding stack on tasks like real repository-level pull requests, have not been published by a neutral party. The Register noted that Meta is entering a category where the incumbents already have developer habit on their side.