Meta Launches Muse Code AI Agent to Automate Engineering

Meta Platforms Incorporated debuted Muse Code, a terminal-based AI coding agent for macOS and Linux, alongside a coding-centric update to its AI model, Muse Spark 1.2. Developed by Meta Superintelligence Labs, the tool automates software engineering tasks including planning changes, writing code, and validating results across large repositories. The agent utilizes persistent asynchronous background agents and a local event log to reduce latency and ensure sessions are restart-safe.

To compete with OpenAI's Codex and Anthropic's Claude Code, Meta is employing a low-pricing strategy. The service costs $1.25 per million input tokens and $4.25 per million output tokens, with a heavily discounted contributor tier for users who allow Meta to use their data for model improvement. While Meta reports that Muse Spark 1.2 outperformed GPT-5.6 Terra on Terminal-Bench 2.1, the model trailed rivals like Claude Opus 5 on DeepSWE 1.1.

Beyond the tool's release, Meta has implemented the assistant as the primary support path for developers, replacing traditional support tickets. This shift aligns with a broader internal strategy to replace various roles, including 20% of staff, with AI agents. This transition follows a June incident where a rogue AI agent inadvertently allowed unauthorized access to over 20,000 Instagram accounts, raising concerns regarding the reliability of AI in handling sensitive developer information.