Meta has fired its most serious shot yet in the AI coding wars. On Wednesday, Meta Superintelligence Labs released Muse Code — a terminal-based AI coding agent now in beta — alongside Muse Spark 1.2, a coding-focused update to its frontier model family. Together they put Meta head-to-head with Anthropic's Claude Code, OpenAI's Codex and Google's Gemini CLI in the fastest-growing corner of enterprise AI.
"Releasing Muse Code in beta today," CEO Mark Zuckerberg wrote on X. "It's a terminal coding agent that takes on complete software engineering tasks across large repos: planning changes, writing code, validating the results."
The headline architectural bet is what Meta calls async background agents. Instead of spawning helper agents fresh for each task, Muse Code keeps specialized agents alive for the whole session, so they don't re-explore a repository every time. For big jobs, it fans out to parallel sub-agents, each working in an isolated git worktree — "your working copy is never touched," Zuckerberg said, claiming tests had it build six features for a game simultaneously "with no collisions."
Every model call, tool run, approval and edit is appended to a local event log before execution, which Meta says makes the runtime "replay-exact and restart-safe": crash 20 hours into a task and it resumes precisely where it stopped.
The model underneath, Muse Spark 1.2, was co-trained with the harness itself using rejection-sampled agent trajectories — a sign of the industry-wide shift toward tuning models for the tools they run inside. On Terminal-Bench 2.1 it scored 82.9%, edging OpenAI's GPT-5.6 Terra in Codex (81.8%) but trailing Claude Opus 5 (86.7%). On Meta's own internal benchmark it reached 70.6%, nearly nine points behind Opus 5's 79.4%.
Pricing is where Meta departs from its rivals. Standard access costs $1.25 per million input tokens and $4.25 per million output tokens, with a commitment that prompts are not used for training. But the "contributor" tier charges $0.10 and $0.20 respectively — roughly 21x cheaper on output — in exchange for explicit permission to train future models on your code and prompts. It is the cheapest coding model on the market, yet the default on-ramp for Muse Code sends developers' code into Meta's training pipeline; enterprises with proprietary codebases must consciously opt out by switching to standard pricing.
Meta AI chief Alexandr Wang framed the cost argument directly: "We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective."
The launch also closes a chapter: Muse Code and Muse Spark 1.2 are fully proprietary, with no downloadable weights, and Zuckerberg's only answer when asked whether Muse would be open-sourced was "I'll have more to share on that soon."
