Meta Launches Muse Code: A New AI Coding Agent Powered by Muse Spark 1.2

Meta Superintelligence Labs has launched Muse Code in beta, a terminal-based coding agent powered by its new Muse Spark 1.2 model, marking a significant strategic entry into the rapidly evolving AI-driven software development landscape. This move directly positions Meta to compete with established players like Anthropic's Claude Code and OpenAI's Codex, particularly by emphasizing a competitive pricing strategy. Muse Code, installable with a single command on macOS or Linux, is designed to handle complete software engineering tasks across large repositories, including planning changes, writing code, and validating results.

The core innovation within Muse Code lies in its architectural approach, featuring "async background agents" that persist throughout a user's session, unlike most rival harnesses that spawn agents per task. Meta claims this design reduces redundant information gathering and lowers latency on multi-step tasks by allowing agents to carry out next steps autonomously and report back only when necessary. Furthermore, Muse Code utilizes an append-only event log as a single source of truth, ensuring runtime replay-exactness and restart-safety, which means the agent can resume precisely where it left off after a crash, crucial for long-running tasks. The agent comes with three bundled skills: `/plan` for approval-gated task planning, `/grill` for stress-testing plans, and `/goal` for objective completion.

Muse Spark 1.2, the multimodal model underpinning Muse Code, was co-trained specifically with the agent's harness, a departure from competitors that often adapt general-purpose foundation models for coding tasks. This co-training approach, which included rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, aims to maximize compatibility and performance when the model and agent are paired. Meta claims Muse Spark 1.2 performs on par with or even surpasses rivals like GPT 5.6 Terra in the Terminal-Bench 2.1 software engineering benchmark and closely approaches Claude Opus 5, though it lags behind both in the long-horizon DeepSWE 1.1 test. The model also boasts multimodal capabilities, processing both text and image input to generate text output, and employs extended thinking or chain-of-thought reasoning for complex problems.

This release is a critical step for Meta Superintelligence Labs (MSL), a division established in June 2025 with the ambitious goal of developing AI that surpasses human intelligence, often referred to as "personal superintelligence". Led by Alexandr Wang, former Scale AI CEO, and Nat Friedman, former GitHub CEO, MSL has been aggressively recruiting top AI talent with substantial compensation packages. The launch of Muse Code directly addresses investor pressure on Meta to monetize its significant AI investments, especially after a recent stock decline. By offering a standard pay-as-you-go option at $1.25 per million input tokens and $4.25 per million output tokens, and a "contributor tier" that is more than ten times cheaper, Meta is strategically undercutting rivals like Anthropic's Claude Code and OpenAI's Codex on price. Meta also supports zero data retention requests for enterprise clients, a key consideration for businesses with sensitive codebases.

The impact of Muse Code on users and the industry could be substantial. The AI coding agent market is rapidly maturing, with tools like Claude Code, OpenAI's Codex, GitHub Copilot, Cursor, and Devin already prominent. While AI coding tools have been shown to increase code volume by roughly 180%, studies indicate that the amount of code actually shipped to production only rises by about 30%, highlighting a "productivity paradox" where immediate task-speed gains can hide downstream costs like increased rework and incidents. Muse Code's persistent background agents and robust event logging aim to mitigate some of these downstream issues by building continuous context and ensuring task resilience, potentially improving the reliability of AI-generated work over longer development cycles. The ability to deploy separate sub-agents for larger tasks in isolated worktrees, as demonstrated by Meta's testing where the system built six game features simultaneously without conflicts, suggests a path toward more complex, parallel software development.

Looking ahead, the competitive landscape will intensify. While Meta acknowledges it is not competing on the "frontier of capability" against the likes of Claude Code and Codex, its aggressive pricing and specialized co-training approach for Muse Spark 1.2 could capture a significant market share, especially among developers and enterprises sensitive to cost and data privacy. The ongoing trend in 2026 sees AI agents moving from mere autocomplete functionality to full-fledged collaborators that plan, write, test, and debug entire features. The success of Muse Code will depend on its ability to bridge the gap between perceived speed and actual productivity, offering genuine gains in shipped software rather than just generated code volume. With Meta's commitment to further harness features and more powerful models on the horizon, the Muse Spark line could evolve to close any existing capability gaps, solidifying Meta's position as a major player in the agentic coding revolution. The industry is shifting towards multi-tool environments where developers combine several AI agents, and Muse Code's terminal-first approach and compatibility with third-party platforms could make it a valuable component in such integrated workflows.