Muse Code is built to help developers with complex programming tasks.
Meta’s HQ in Cupertino, USA. Picture: X/@EliBurton_
Meta has unveiled Muse Code, its first coding agent, now in beta and available worldwide through the Meta Model API.
The launch comes only a month after Muse Spark 1.1, showing how quickly the company is moving in AI development.
Muse Code
Muse Code is built to help developers with complex programming tasks.
Instead of working file by file, it can manage entire projects. Users give plain‑English instructions, and the agent plans the work, writes the code, tests it, and checks the results.
It can also run several tasks at once, making it faster than traditional tools.
The agent is powered by Muse Spark 1.2, Meta’s newest model. This version was trained alongside Muse Code, so the two work closely together.
Context window
It also has a massive one‑million‑token context window, which means it can keep track of large codebases without losing its place.
“We co‑trained Muse Spark 1.2 with Muse Code to ensure the model exhibits its best performance and coding usability when paired together,” Meta said.
Different approach
While Meta is pushing ahead with Muse Code, other tech companies are taking different approaches.
Salesforce says AI coding agents have already boosted developer productivity by more than 50%, allowing smaller teams to deliver more software without hiring.
By contrast, Discovery’s CIO Derek Wilcocks cautions that the technology is still limited, warning that explainability and maintainability remain unresolved and that human oversight will remain critical.
Reliability
Reliability is another focus for Meta. Every step is logged, so if the system crashes, Muse Code can resume exactly where it stopped.
“That ability lets Muse Code take on long‑running tasks without being derailed by failures,” the company explained.
Pricing
Pricing comes in two tiers: a standard rate and a cheaper “contributor” option for those who opt in.
Early tests show Muse Code can handle demanding jobs, such as improving GPU kernel performance over thousands of tool calls.
Meta believes this is only the beginning, saying it marks the next step toward the frontier, with larger and much more capable models on the way.
