Meta Unveils Muse Code: A Coding Agent That Rivals Opus 5

Meta officially released the Muse Code beta, the company’s first AI coding agent that runs in a terminal and is driven by the Muse Spark 1.2 model. Mark Zuckerberg said it can perform complex software‑engineering tasks in large repositories, including project analysis, planning, code generation, tool execution, and result verification.

Muse Code supports macOS and Linux and can be installed with a single terminal command. Users can give a complete requirement—such as fixing a multi‑module bug, adding a new feature, or refactoring a large project. The agent first understands the codebase, then creates a plan, modifies files, runs tests, and iteratively adjusts based on the outcomes.

The system uses a simple agent loop together with a set of asynchronous background agents that stay alive for the entire session. These background agents continuously run, reducing repeated information gathering, can execute subsequent steps autonomously, and decide when to feed results back to the main agent, thereby lowering latency and dependence on human guidance.

Runtime design relies on a local event log that records every model call, tool execution, approval action, and code edit. This log serves as the single source of truth, enabling precise replay, safe restart after crashes, and continuation from the interruption point.

Built‑in skills are exposed via commands such as /plan (breaks a task into an approval‑required plan), /grill (stress‑tests the plan until it is reliable), and /goal (drives progress toward a specified goal until the task succeeds).

Muse Spark 1.2 is an upgrade over 1.1, focusing on code generation, complex debugging, code‑base understanding, and end‑to‑end development performance. Training added substantial compute for programming tasks and diversified the training environments while preserving strong general‑agent abilities.

Benchmark results show Muse Spark 1.2 trailing only Opus 5 (max) on Terminal‑Bench 2.1. On DeepSWE 1.1—113 tasks across 91 repositories in TypeScript, Go, Python, JavaScript, and Rust—it is weaker than Opus 5 and GPT‑5.6 Terra. On Meta Internal Coding Bench (440 real pull‑request tasks) it also falls behind Opus 5.

The strong scores stem from three technical advances:

Co‑training with Muse Code: Muse Spark 1.2 and Muse Code are jointly trained so they work together efficiently. Training introduced rejection‑sampled agent trajectories, optimized execution, context compression, and sub‑agent handling, and incorporated Muse Code’s toolset.

Long‑duration task capability: Large‑scale training on full‑repo generation, end‑to‑end projects, and automated research. The model plans task order, uses goal‑condition constraints to stay on target, and applies context compression to retain essential information.

Self‑improvement: Muse Spark 1.1 generated difficult programming environments and instruction‑following templates; a model evaluated candidate solutions, creating scalable training data for 1.2, which improves instruction‑following accuracy.

Pricing is $1.25 per million input tokens, $0.15 per million cached input, and $4.25 per million output tokens for the standard tier; the Contributor tier costs $0.10 per million input, $0.002 cached input, and $0.20 output.