Key facts
- Meta launched Muse Code (beta), an AI coding agent powered by its Muse Spark 1.2 model.
- The agent features a crash-safe runtime with an event log for exact resumption of tasks.
- Muse Code can write code, verify results, and handle complex coding projects by running multiple sub-agents.
- Meta's internal benchmarks show Muse Spark 1.2 scoring 82.9% on Terminal-Bench 2.1, trailing Anthropic's Opus 5 but ahead of OpenAI's Codex.
- The tool is available via the Meta Model API and a curl install script on a pay-as-you-go plan.
Meta Platforms has launched Muse Code (beta), a new terminal coding agent powered by its updated Muse Spark 1.2 model. The agent is designed to handle complex software engineering tasks across large code repositories, coordinating multiple persistent subagents to improve speed, accuracy, and reduce intervention.
A standout feature of Muse Code is its runtime logging, which records every model call, tool run, approval, and edit to a local event log. This ensures that the agent can resume precisely where it left off after a crash, a critical capability for long-running jobs that competitors have not emphasized.
Meta's internal benchmarks show Muse Spark 1.2 scoring 82.9% on Terminal-Bench 2.1, trailing Anthropic's Opus 5 at 86.7% but surpassing OpenAI's Codex (81.8%) and Google's Antigravity (81.6%). On the DeepSWE 1.1 benchmark for agentic coding capabilities, Muse achieved 59.3%, compared to Opus 5's 65.0% and Codex's 64.8%. Meta's internal coding benchmark placed Muse at 70.6% against Opus 5's 79.4%.
While Muse Spark 1.2 may not lead in raw benchmark scores, Meta highlights its performance in speedups over accumulating tool calls, showing improvements of roughly 61–69%. The company also demonstrated multimodal capabilities, where Muse Code interpreted a video of a house to generate a visually rich website with booking features.
Meta is entering a competitive landscape, with established agents like OpenAI's Codex and tools from Anthropic already in the market. Muse Code's primary advantages are its robust crash-safe runtime and sophisticated subagent design, rather than outright benchmark superiority. The company acknowledges the inherent risks of powerful, autonomous agents but is betting on developer demand for such capabilities.