Three months after releasing its first Muse AI model, Meta is rolling out a significant update as it tries to close the gap with market leaders. The new model, Muse Spark 1.1, is Meta's strongest tool yet for coding and autonomous AI agents, according to Alexandr Wang, the company's AI chief. The original Muse Spark, launched in April, was limited to a small group of partners through a private API preview. Now, Meta is opening access to developers via a public preview on its own platform, with a waitlist for new users.

Meta is positioning the model as a direct competitor to offerings from OpenAI and Anthropic, particularly in pricing. Wang described the rates as "very aggressive and attractive": new accounts get $20 in free credits, then pay $1.25 per million input tokens and $4.25 per million output tokens. He said Muse Spark 1.1 outperforms rival models in tasks that involve integrating with third-party coding tools and services. The move comes as Meta faces pressure from investors to show returns on its massive AI infrastructure spending, which rivals that of cloud giants—though Meta lacks a major cloud business of its own.

The model was trained by Wang's Meta Superintelligence Labs to excel at coding, which Wang says is foundational for building AI agents capable of handling complex, multi-step tasks. Meta is also releasing Muse Image, a model for generating pictures, aimed at creators and advertisers. While Meta previously focused on open-sourcing its Llama models, it is now selling access to proprietary AI. Wang said a future version of Muse Spark will be open-sourced, but didn't say when. Meta is already training a more powerful model, code-named Watermelon, with no release date set.

Source: CNBC