Meta released Muse Spark on Tuesday, the first AI model to emerge from the company's Superintelligence Lab — and the results are mixed.

According to the New York Times, Muse Spark outperforms Meta's previous generation of AI models across several benchmarks. But it stumbles on coding tasks, an area where rivals like OpenAI's GPT-5 and Anthropic's Claude have pulled ahead.

That's a problem. Coding ability has become the litmus test for advanced AI systems. Developers use these models to generate software, debug errors, and automate repetitive tasks. If your model can't code well, you're selling a sports car with a missing wheel.

The Superintelligence Bet

Meta launched its Superintelligence Lab with fanfare last year, positioning it as the company's moonshot toward artificial general intelligence. The lab operates separately from Meta's main AI research division, with its own compute resources and talent pool.

Muse Spark is the first tangible output. It's a test of whether Meta's organizational restructuring can accelerate innovation — or whether it's just shuffling deck chairs.

The coding gap suggests Meta still has ground to cover. OpenAI and Anthropic have spent years refining their models on massive code repositories, building partnerships with developer tools, and iterating based on real-world usage. Meta is playing catch-up.

The question now: Can the Superintelligence Lab iterate fast enough to matter? Or will Muse Spark become another footnote in the AI arms race, outpaced before it finds a market?

Meta hasn't disclosed pricing or availability details yet. For now, Muse Spark exists in that awkward space between announcement and relevance.