Meta just made its play for the AI coding agent wars. The company has officially launched Muse Code in beta, moving the tool into direct competition with Anthropic’s Claude Code and OpenAI’s Codex.
Powered by the Muse Spark 1.2 model, Muse Code is a terminal-based AI coding agent built for complex software engineering tasks. It handles planning, code modifications, testing, and multi-agent coordination through persistent background sub-agents, all from the command line.
Pricing designed to undercut the competition
The standard tier charges $1.25 per million input tokens and $4.25 per million output tokens through the Meta Model API. A “contributor” tier is priced roughly 10 to 20 times cheaper than the standard option. The catch: you grant Meta permission to use your usage data for model improvements.
The API itself is designed to be compatible with existing SDKs from OpenAI and Anthropic. Installation is available via command-line for macOS and Linux users.
Benchmarks paint a competitive picture
On Terminal-Bench 2.1, Muse Code scored 82.9%, placing it second behind Claude Opus 5 and Claude Code, which achieved 86.7%. On the DeepSWE 1.1 benchmark, it ranked third overall with a 59.3% score.
Muse Code supports multi-agent orchestration, spinning up persistent sub-agents that work in the background on different parts of a problem simultaneously. It uses isolated Git worktrees for parallel work, so multiple agents can modify code without stepping on each other’s changes. For reliability, it features crash-recovery through an append-only local event log. Meta highlighted an internal study where the tool handled over 1,000 tool calls across a continuous 24-hour period.
The AI coding arms race heats up
The development is led by Alexandr Wang at Meta Superintelligence Labs. The Muse Spark model family first appeared in April 2026, with Muse Spark 1.1 arriving in July 2026. Muse Code is built on Muse Spark 1.2.