Meta Releases Muse Code Agent and Muse Spark 1.2

Meta has introduced its inaugural software engineering agent, Muse Code, in a beta release designed to handle complex programming workflows. The terminal-based tool installs via a single command and manages entire development cycles, from planning and writing code to validating the final output. To achieve this, Muse Code coordinates multiple specialized agents, deploying parallel workers to implement code changes while background reviewers continuously monitor the process.

Powering this agent is Muse Spark 1.2, a newly updated flagship model that Meta co-trained with Muse Code to minimize retries and optimize tool usage. The model features a 1 million-token context window, allowing developers to execute long-running engineering tasks within a single session. Meta has highlighted the model's performance on several benchmarks, including Terminal-Bench, DeepSWE, GDPval, and its own Meta Internal Coding Bench, though the company has not yet published a detailed methodology for these evaluations.

Developers can access the new model through the Meta Model API, which has entered public preview with expanded global availability, or via OpenRouter. Meta offers two distinct pricing structures. The standard tier costs $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens, with a guarantee that user prompts will not be used to train Meta's products.

For teams willing to share their data to help improve Meta's models, the contributor tier significantly lowers these costs. Under this option, pricing drops to $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens. This tiered approach allows developers to choose between strict data privacy and highly subsidized operational costs.