Particle.news

Meta Launches Muse Code Beta and Muse Spark 1.2

The release signals Meta's bid to win long‑running engineering work by pairing a co‑trained coding model with a low‑cost contributor tier that lets the company use developer prompts to train future models.

Overview

  • Meta announced Muse Code in beta on Wednesday, August 5, 2026, a terminal‑based coding agent for macOS and Linux that installs with a one‑line command and runs locally while coordinating persistent subagents.
  • Muse Spark 1.2 was co‑trained with Muse Code to improve code generation, complex debugging, and whole‑repository workflows and the system logs every model call and tool action so long jobs can resume exactly after a crash.
  • Meta demonstrated long‑horizon capability with a GPU kernel optimization that ran more than 1,000 tool calls over runs up to 24 hours to show the agent can iterate and improve across extended engineering tasks.
  • Pricing is pay‑as‑you‑go at $1.25 per million input tokens and $4.25 per million output tokens for the standard tier, plus a deeply discounted contributor tier that charges about $0.10/$0.20 per million in exchange for allowing Meta to use prompts and completions for training.
  • Meta is feeding internal engineer corrections from its MetaCode program into training and positions Muse Code to undercut Anthropic and OpenAI on cost even as published benchmarks show Muse Spark 1.2 competitive on some tests but trailing Anthropic's Opus 5 on several coding benchmarks.