Meta Platforms has delayed plans to release its newest artificial intelligence model to developers multiple times and as of Tuesday didn’t have a planned date to release it, according to people familiar with the matter.
The delay, stretching nearly two months after the company’s AI chief told developers to expect a release “soon,” raises questions about how quickly Meta can monetize its massive investments in building its own frontier AI models.
The company has been developing an application programming interface, or API, a software tool that allows different programs to talk to each other. Meta’s API would allow apps written for computers or mobile phones to be based on Meta’s AI technology.
In response to questions from The Wall Street Journal on Wednesday, a Meta spokesman said the company was testing the API with partners and planned to release it this month. “We know people want the API and we’re excited to get it into their hands,” the spokesman said.
For so-called “closed” AI models, meaning models that are kept proprietary and not publicly available for download, the only way for developers to access them is via API. Companies typically release an API alongside a new model, or within a few weeks of releasing it, to maximize its relevance among developers.
OpenAI and Anthropic make money in part by selling access to their APIs, which customers use to embed their AI into custom projects and tools without having to build their own models from scratch.
Meta initially planned to release the API around the time it launched its latest AI model, called Muse Spark, in April, some of the people said. Two days after its release, Meta’s chief AI officer, Alexandr Wang, posted on X that the API would be launching shortly.
“The muse spark API will be coming soon! we have been thrilled with the amount of excitement amongst developers who want to try muse spark inside their agentic harnesses,” he wrote. “Stay tuned!”
The API never came out. The first delay, from April to May, was attributed to bugs that became apparent in testing and the need to build more infrastructure, according to people familiar with the matter. It was then delayed again to June, one of the people said.
Meta plans up to $145 billion this year in capital expenditures, largely to build out its AI infrastructure efforts. The company has lofty ambitions of creating personal and business agents for its 3.5 billion daily active users and is building AI models to achieve those goals.
In April, Wall Street balked at Meta’s plans to spend even more money this year than previously anticipated, sending its stock down more than 5% in after-hours trading.
Meta has started sharing its playbook for recouping some of those costs. Last week, the company announced new subscriptions for Instagram, WhatsApp and Facebook and said it would start testing subscriptions to its AI chatbot, Meta AI.
In a separate call with shareholders, Meta Chief Executive Mark Zuckerberg said creating a cloud computing business was “definitely on the table” as a way to monetize any excess capacity it builds.
Zuckerberg also said companies come to Meta every week asking it to establish an API service. But he did not give an update on when Meta might launch the Muse Spark API.
Meta delayed the launch of one of its AI models last year after company engineers struggled to significantly improve its capabilities, The Wall Street Journal previously reported. That model, called Behemoth, was never ultimately released.
Instead, Meta went on a hiring blitz and overhauled its AI teams, picking Wang as chief of the newly formed Meta Superintelligence Labs. One division of MSL, the secretive unit called TBD Lab, developed the model that would become Muse Spark.
The AI models Meta released before Wang’s arrival were all open-source, meaning developers could freely download and use them. Muse Spark, which powers Meta’s AI chatbot and AI features within it, is the first to be released without sharing the blueprint and software files behind it.
According to Meta’s own internal benchmark evaluations, the model was competitive with OpenAI and Anthropic and significantly outscored xAI’s Grok on most tests. But developers largely haven’t been able to access the model themselves beyond a few third-party evaluation firms that were given access to run their tests on Meta and provide scores ahead of its launch.
Write to Meghan Bobrowsky at meghan.bobrowsky@wsj.com