Meta has launched Muse Code, a terminal-based beta coding agent for macOS and Linux that uses persistent background agents to handle complex, multi-step software tasks across large codebases. It runs on the new Muse Spark 1.2 model, which was co-trained alongside the agent to improve planning and context handling. On Terminal-Bench 2.1, Muse Spark 1.2 scored 82.9% (behind Claude Opus 5, slightly ahead of GPT-5.6 Terra), and 59.3% on DeepSWE 1.1, trailing both rivals. Analysts note the co-training approach is not unique since OpenAI and Anthropic follow similar strategies, and meaningful competitive advantage must be proven on real enterprise pipelines. Enterprise adoption faces hurdles including CI/CD integration resistance, data governance concerns, vendor lock-in fears, and unpredictable token costs. Meta's two-tier pricing model also raises data-governance tradeoffs, with the cheaper Contributor tier potentially used to improve Meta's products.
