Alexander Wang, Chief AI Officer at Meta, at the Bloomberg Tech Conference on Thursday, June 4, 2026 in San Francisco, California, USA.
David Paul Morris | Bloomberg | Getty Images
Three months after announcing its first artificial intelligence model under the leadership of AI chief Alexander Wang; Meta is rolling out major updates to compete with OpenAI and Anthropic in key areas of the market.
Muse Spark 1.1, which Meta introduced Thursday, is the company’s “most powerful model to date for agent and coding work,” Wang said in an interview with CNBC. The initial Muse Spark model, released in April, was only available to “select partners” who could access the technology through a “private API preview.”
Meta is making the new model API available through its developer portal as part of a public preview. Users can sign up there and review the integration instructions. A Meta spokesperson said some early partners already have access to the API, and new users can “add themselves to a waiting list and be added from there over time.” For now, Meta said it is restricting API access to its properties rather than making the API available on third-party platforms like the popular OpenRouter marketplace.
“This will be delivered on top of the computer infrastructure that we have built,” Wang said.
Meta’s second notable development for the Muse family this week. Meta Inc. on Tuesday released Muse Image (originally codenamed Mango), a model for image creation, aiming to attract creators and advertisers to its products.
Meta CEO Mark Zuckerberg is under pressure from Wall Street to show a return on the company’s huge and growing investments in AI infrastructure and development. Meta is spending at the same pace as its hyperscaler peers, but it doesn’t have a cloud infrastructure business (though it has plans to start one), and has a large number of companies including OpenAI, Anthropic, google For developing popular models and AI applications.
Wang characterized the pricing for the Muse Spark update as “very aggressive and attractive” compared to similar services offered by labs like Anthropic and OpenAI. He said all new API accounts start with $20 in free credit. From there, the company plans to charge $1.25 per million input tokens and $4.25 per million output tokens, he said.
“The goal is to achieve attractive pricing for huge consumption volumes,” Wang said.
He said Muse Spark 1.1 performed better than rival models in certain tasks, including the ability to interact with various third-party coding products and services.
Wang’s Meta Superintelligence Lab (MSL) trained Muse Spark 1.1 to excel at coding-related tasks. He said this ultimately improves the AI agent’s ability to perform multiple tasks autonomously, like a troupe of human interns.
“We need to build coding capabilities as part of our overall agent service offering,” Wang says.
The technology industry’s excitement about AI agents increased in the first half of 2026. This is in part due to the sudden popularity of OpenClaw, which developers can use to manage the AI models that power powerful digital assistants. Wang said Meta trained Muse Spark 1.1 to “work well with all of the most common harnesses that developers are using today. Given our goal of maximizing adoption, we felt that was the best approach for this model.”
While Meta’s previous AI strategy focused on releasing early Llama family models to the open source community, the company is now focused on selling access to its own AI models.
Wang said Meta is still “committed to open source” and that his MSL division has “a variant of Muse Spark that we are developing with the intention of open sourcing.” He declined to say when the company would release it.
Wang added that he is “dogfooding” the latest Muse Spark model and is excited that the technology can be used as a tool to improve personal health through tasks such as searching the web, reading academic papers, and accessing personal health-related data.
“This is one of the use cases that I think really encapsulates the need for these agent systems,” Wang said of the AI and health experiments.
Wang said Meta is currently training a more powerful AI model, code-named “Watermelon,” but did not say when it would be released. Muse Spark’s code name was Avocado.
Attention: One year since joining the Meta Superintelligence Institute.
