Meta has unveiled a public preview of its new model, Muse Spark 1.1, focused on coding and AI agent tasks.

CNBC reported on July 9 that Meta released the API through its developer portal and has entered a coding AI race with OpenAI and Anthropic.

The update comes three months after Alexandre Wang (알렉산드르 왕) took charge of AI and Meta unveiled its first AI model.

Wang stressed that Muse Spark 1.1 is Meta's most powerful model for agent and coding tasks. Meta's first Muse Spark, released in April, was provided only to a select group of partners in a private API preview.

Meta is running the new model API as a public preview on its developer portal. Some early partners can already use the API. New users receive access sequentially after joining a waitlist. The API is currently available only through Meta's own service, not external platforms such as OpenRouter.

Meta also set low pricing. New API accounts receive $20 in free credits. After that, Meta charges $1.25 per 1 million input tokens and $4.25 per 1 million output tokens.

Meta has continued to roll out models in the Muse family this week. On July 8 it unveiled Muse Image, an image-generation model that had the codename Mango. Meta is seeking to draw creators and advertisers into its own AI product lineup with the model.

Meta stressed that Muse Spark 1.1 outperformed rival models on some tasks. Wang said it showed strength in tasks involving a range of external coding products and services.

Meta's AI strategy is also changing. Meta put weight on releasing its Llama family of models to the open-source community, but is now focusing on selling access to closed AI models. Meta is also developing a Muse Spark variant model with openness in mind. It did not disclose a release timing.

Meta is also training more powerful AI models through a project with the codename Watermelon. It did not disclose a release schedule.