Meta enters the terminal arena with Muse Code AI agent

Meta officially rolls out Muse Code, an AI agent designed to operate natively within the terminal. Developed under the umbrella of Meta Superintelligence Labs (MSL), this ambitious project aims to streamline developer workflows, offering a direct response to established competitors like Anthropic and OpenAI. Powered by the newly unveiled Muse Spark 1.2, which succeeds the earlier 1.1 iteration, the agent focuses on providing end-to-end management of complex programming tasks, attempting to overcome the persistent hurdles associated with long-horizon coding workflows.

By opting for a terminal-first deployment that requires a simple command line installation, Meta aims to eliminate friction for software engineers. Instead of forcing developers into external interfaces or clunky web dashboards, Muse Code fits right into existing environments. What truly sets the system apart, however, is its architectural approach to handling massive projects. Through a feature termed Agent Fan-out, the central agent can decompose large, multi-faceted prompts into smaller sub-tasks. It then dispatches specialized sub-agents to operate simultaneously across isolated git worktrees. This parallel processing design ensures that experimental changes or feature additions occur in sandbox environments without corrupting the main active working tree or triggering code conflicts.

Underpinning this parallel orchestration is a context window capable of processing 1 million tokens, allowing the underlying model to analyze complete codebases rather than isolated fragments. This comprehensive repository understanding enables the agent to respect existing architectural conventions while generating or refactoring code. Furthermore, Meta has integrated a persistent, transparent local event log that records every action, decision, and API interaction. Beyond providing full auditability for engineering leads, this event tracking acts as a crash recovery safeguard. If a long-running procedure gets interrupted, developers can resume the session instantly without starting over or re-querying the model.

Perhaps the most talked-about aspect of this rollout is Meta's aggressive dual-tier pricing model, which highlights the ongoing industry scramble for training data. Users can opt for the Contributor tier at an extremely low price of 0.10 dollars per million input tokens, provided they grant Meta permission to utilize their code interactions for future model training. For enterprise clients, professional studios, or developers bound by strict non-disclosure agreements, the Standard tier charges approximately 1.25 dollars per million input tokens. In exchange for this higher rate, Meta guarantees strict data privacy, assuring commercial teams that their proprietary code will never be harvested to train future iterations of Llama or Muse Spark models.

On the benchmark front, Meta has taken a transparent approach regarding performance. Evaluated on the rigorous Terminal-Bench 2.1 benchmark, Muse Spark 1.2 earned an impressive score of 82.9 percent, edging out OpenAI Codex running on GPT-5.6 Terra at 81.8 percent. However, it still trails Anthropic's market-leading Claude Code running on Opus 5, which tops the benchmark at 86.7 percent. Despite sitting in second place on raw benchmark metrics, Meta argues that the orchestration capabilities of Muse Code offer distinct practical advantages. The ability to deploy multiple sub-agents simultaneously often yields faster real-world execution times for debugging and refactoring, proving that architectural execution can be just as crucial as raw model intelligence in daily software development.