Meta debuts Muse Code agent to challenge OpenAI and Anthropic

Muse Code is Meta’s declaration that the coding‑agent race is now a three‑way fight. Rather than just shipping another GitHub Copilot clone, Meta is betting on a more ambitious pattern: persistent agents that live in your terminal, span an entire repository and survive crashes or restarts by replaying an append‑only event log. Combined with a long‑context Muse Spark 1.2 model, that shifts agents from “autocomplete that runs tests” toward something closer to a junior engineer that can own multi‑day refactors.

Strategically, this hits three nerves. First, it directly targets high‑value enterprise spend where OpenAI’s code models and Anthropic’s Claude Code have been gaining traction. Second, the pricing mix, including a steeply discounted “Contributor” tier, is clearly designed to pull in data and usage at scale, echoing Meta’s playbook with Llama. Third, by integrating the harness and the model training, Meta is signaling that agent performance will be optimized end‑to‑end, not just at the model level. If Muse Code proves reliable on large commercial codebases, it will pressure rivals to match both persistence and price.