Meta has shipped Muse Code, its first coding agent, in a beta launch alongside Muse Spark 1.2, a brand new model of its flagship mannequin that the corporate says was co-trained with the agent for tighter integration.
Muse Code is a terminal agent that installs with a single command and takes on full software program engineering duties, together with planning modifications, writing code, and validating outcomes. On its Muse Code product web page, Meta describes it as “an agent to your most advanced coding workstreams,” with a number of brokers coordinating on every job: parallel staff doing the implementation whereas reviewers run within the background. The corporate says each motion the agent takes is clear and traceable, and that as a result of Muse Code and Muse Spark 1.2 have been co-trained, the pairing produces higher instrument use, fewer retries, and higher-quality output than a generic wrapper round an outdoor mannequin.
The underlying mannequin, Muse Spark 1.2, is optimized for actual coding workflows with what Meta describes as increased first-attempt accuracy and extra dependable instrument calling. It carries a 1 million-token context window, which Meta says lets long-running duties run begin to end in a single session. Meta publishes vendor-reported benchmark charts for the mannequin on the web page, together with outcomes on Terminal-Bench, DeepSWE, its personal Meta Inside Coding Bench, and GDPval, although the web page presents them as pictures and not using a methodology write-up.
Pricing splits right into a contributor tier and an ordinary tier
Entry runs via the Meta Mannequin API, which Meta says is now in public preview with expanded international entry. The usual Muse Spark 1.2 tier is priced at $1.25 per million enter tokens, $0.15 per million cached enter tokens, and $4.25 per million output tokens, and prompts on that tier should not used to enhance Meta’s merchandise. A contributor tier drops the value to $0.10 per million enter tokens, $0.002 per million cached enter tokens, and $0.20 per million output tokens, in trade for letting Meta use the info to enhance its fashions. The mannequin can also be out there via OpenRouter for builders who wish to swap it into present instruments.
