Meta Releases More Powerful AI Model, Edging Closer to Rivals

Meta released its most powerful artificial intelligence model yet, with its chief AI officer saying its capabilities are edging closer to top competitors.

Developers on Wednesday will be able to access and pay for Muse Spark 1.3, an updated version of its most advanced model, the company said. It will soon roll out the update to users of Meta’s social media platforms, like Instagram and Facebook, as well as Meta AI.

“This is our biggest jump so far on model performance,” Meta’s Chief AI Officer Alexandr Wang said in an interview on Wednesday, pointing to advancements in the model’s coding and agentic capabilities, or ability to complete tasks on behalf of human users. Wang said the update put Meta on par with recently released AI models from its rivals OpenAI and Anthropic.

Wang said Muse Spark 1.3 is “competitive” with Anthropic’s model, Claude Fable 5.1, and “better than” OpenAI’s GPT-5.6 Sol, particularly when it comes to coding, though OpenAI will soon be releasing a new, even more advanced model known as Astra. In his opinion, Meta’s model outperforms “any of the current Chinese models out there.”

It can be difficult to compare AI models, as some are better at some tasks and worse at others, and benchmarking parameters can be gamed and aren’t necessarily reflective of how the technology performs in the real world.

Chief Executive Officer Mark Zuckerberg recently wrote a public note explaining the importance of making AI development more open and accessible. But Meta needs to balance those goals against returns on its massive investment in AI infrastructure and talent.

Wang said Meta has not yet decided whether or not it will release the Muse Spark 1.3’s weights — the internal blueprint, formed during the training process, that determines how the model responds. Releasing the weights would allow outside developers to download, run or build upon the model on their own. The company still plans to release weights for the prior version, Muse Spark 1.2, Wang said.

The Meta AI chief said the company remains on track for the development of its larger, highly anticipated model dubbed Watermelon.

“We believe that Watermelon will be extremely competitive,” he said, though he declined to provide specific timing for its release.

Meta is spending hundreds of billions of dollars in a bid to catch up in a fast-moving AI race. Last year, Zuckerberg overhauled the company’s strategy, poached Wang from a company he founded, and put him at the helm of a new unit called Meta Superintelligence Labs, or MSL. Since then, Wang has sought to push out new models and products at a steady clip and close the technological gap with competitors.

But the Menlo Park, California-based company’s aggressive spending spree has drawn the scrutiny of investors looking for a clear return on investment. Amid the skepticism, Meta has begun working on a cloud infrastructure business that will sell access to AI computing power and models, Bloomberg News reported in July.

That same month, Meta for the first time began charging developers for access to its Muse Spark 1.1 model. Zuckerberg said in an interview at the time that he was seeking to make it one of the most affordable options on the market.

Wang said Muse Spark 1.3 would cost developers the same amount as Muse Spark 1.2 through Meta Model API, the platform that providers developer access to the tools. Wang said the API platform has seen strong adoption, with some developers “using trillions of tokens a week.”

Wang said Meta’s latest model has advanced in various ways: It requires 25% fewer tokens, meaning it’s operating more efficiently; it can keep multiple workflows going at once, rather than requiring separate sessions; it’s better at managing long and complex instructions and retaining details across multiple tasks; it has better awareness of its own limitations; and it’s safer — asking for clarification when need and seeking confirmation before taking actions that could be irreversible.

“We did extensive safety testing and safety training across the board for this model before making decision to release it,” he said.

An earlier Meta model accessed the internet and hacked into an outside service’s systems during cybersecurity testing, echoing other recent incidents from model companies that have raised concerns about control over their technology. Wang said the incident helped inform the company’s approach to shoring up safety and security for Muse Spark 1.3.