When Mark Zuckerberg announced he was spending $15 billion to build Superintelligence Labs and poaching Scale AI's 27-year-old CEO Alexandr Wang to run it, the message was clear: Meta was done playing catch-up in AI. It was going straight for the top.

Three weeks later, the lab has shipped its first model. It is called Muse Spark. And by every early measure that matters, it is not the top. It is not even close.

The Benchmarks Are Not Kind

Muse Spark launched as a proprietary model, which was itself a betrayal of Meta's open-source identity. But the bigger problem is performance. Early evaluations show Muse Spark trailing Anthropic's Claude Opus 4.6, OpenAI's GPT-5, and even Google's Gemini 3 across standard benchmarks. For a model that was supposed to represent the beginning of a superintelligence program, being outperformed by competitors who have been in market for months is not a great debut.

Meta's stock rallied on the announcement anyway. Wall Street loves a narrative more than a benchmark. But the developer community, which actually has to use these tools, has been considerably less impressed.

The Open Source Reversal Stings Worse

Meta built its entire AI brand on Llama and open source. "Open source is the path to good AI," Zuckerberg said repeatedly. Llama became the most widely adopted open-weight model in the world. Developers built entire ecosystems around it. Startups raised money on the assumption that Meta would keep the models free.

Then Muse Spark dropped as proprietary. No weights. No downloading. No running it locally. The open-source faithful who spent years evangelizing Meta's approach got nothing. The first model is here and it is not even good enough to justify locking people out.

The Alexandr Wang Bet

Hiring a 27-year-old to run a $15 billion division was either visionary or reckless. Wang comes from Scale AI, which built the data labeling infrastructure that trained most frontier models. He knows the plumbing better than almost anyone. But knowing how to label training data and knowing how to build a model that competes with Anthropic and OpenAI are very different skills.

The charitable read is that Muse Spark is a first draft. The lab has been operational for less than a month. Rome was not built in a day, and neither are frontier AI models. The less charitable read: $15 billion is a lot of money to spend on something that cannot beat a model that costs $20 a month to use.

What This Tells You About the AI Race

Three months ago, there were two serious AI labs: OpenAI and Anthropic. Google was dangerous but distracted. Meta was the open-source wild card. Now the picture looks different. Anthropic just passed OpenAI in revenue. Google is shipping Gemma models that developers actually want. And Meta, despite spending more than anyone, just shipped a model that lost to all of them.

The AI race is not won by who spends the most. It is won by who ships the best model. Right now, Meta is proving that $15 billion and a famous hire can still get you a product that lands with a thud. If Muse Spark 2 does not dramatically close the gap, Zuckerberg will have spent more than the GDP of some countries on an AI lab that made Wall Street rich and developers frustrated.