- Published on
Here’s How It Compares to Claude Code and Codex
- Authors
- Name
- DeepHorus
Meta has released Muse Code, a beta terminal coding agent powered by the new Muse Spark 1.2 model, immediately entering the AI coding agent race. Available now via the Meta Model API and a curl install script, it directly targets developers who rely on tools like Anthropic's Claude Code and OpenAI's Codex.
The launch comes as Meta trails rivals in the agentic AI space, but the company promises bigger models soon. Muse Code handles complex engineering tasks across large repositories, planning changes, writing code, and validating results. It can coordinate multiple persistent subagents for each task, solving problems faster with less intervention.
A key technical feature is a replay-exact event log that records every model call, tool run, approval, and edit. This makes the runtime restart-safe, so a crash resumes exactly where it stopped—a differentiator for long-running jobs. The agent also ships with /plan, /grill, and /goal skills to structure development workflows.
Meta's benchmarks show a mixed picture. On Terminal-Bench 2.1, Muse Spark 1.2 scores 82.9%, behind Claude Code on Opus 5 (86.7%) but ahead of Codex (81.8%). On DeepSWE 1.1, the gap widens: 59.3% vs. 65.0% for Opus 5. Meta admits Opus 5 leads on every chart shown, yet claims Muse beats Codex and Antigravity on most benchmarks.
Long-horizon demos are impressive. Muse Code iteratively optimized GPU kernels for up to 24 hours with over 1,000 tool calls, showing continuous improvement. Another demo showed it converting a home walkthrough video into a functional booking website, highlighting multimodal capabilities.
Developer reactions have been cautiously optimistic, with many praising the replay-exact logging as a novel reliability feature. Analysts argue that real-world performance and trust matter more than benchmark charts, and Meta faces an uphill battle against established competitors. Enterprise demand for AI coding assistants remains high, with open-source options also gaining traction.
Meta says Muse Spark 1.2 was co-trained with Muse Code, ensuring synergy between the model and the agent. Larger, more capable models are already on the way, signaling an aggressive push into agentic infrastructure. The broader industry trend points to AI coding agents becoming indispensable tools, and Meta is determined to claim its share.
