Muse Code: Meta's answer to coding agents from OpenAI and Anthropic

Meta's new programming agent Muse Code uses parallel background agents and the language model Muse Spark 1.2 at comparatively low prices. The beta terminal agent can analyze software projects, plan changes, write code, and check results with tests via the command line. Muse Spark 1.2 was trained jointly with Muse Code and aligned with its tools. Competing with Claude Code from Anthropic and Codex from OpenAI, Meta focuses on low usage costs. Multiple background agents remain active throughout the session, collecting information, handling subtasks, and reporting results to the main agent. A local event log records model calls, tool calls, approvals, and changes, allowing the agent to resume exactly where it stopped after a crash. Meta offers two price tiers: a heavily discounted Contributor plan with limited token quota, available only in selected countries, and a regular usage-based API plan. The independent benchmark platform Artificial Analysis ranks Muse Spark 1.2 among the most cost-effective models in its performance class, scoring 54 points, roughly on par with Grok 4.5 and GPT-5.5, but behind the top models from Anthropic and OpenAI. In coding benchmarks, results are mixed, with the model performing slightly better than its predecessor in one test and slightly worse in another. Artificial Analysis tested Muse Spark 1.2 in its test environments, not Muse Code as a full product. How well the overall package competes with Claude Code and Codex in everyday use remains open. With Muse Spark, developed by Meta Superintelligence Labs, Meta aims to catch up with the most powerful AI systems. Muse Spark 1.2 is the third release after the original Muse Spark and Muse Spark 1.1 in July. Unlike Llama, Meta does not release model weights.