Meta spent years as the company that gave away its AI models for free. That era is now quietly ending.

On September 2, 2026, Meta released Muse Spark 1.3, its most capable AI model to date, and opened API access to developers the following day on a paid basis. The move is a significant signal: the company best known for making its Llama models freely available is now building a proprietary frontier product it intends to monetize directly.

Alexandr Wang, Meta’s chief AI officer, said the model’s capabilities are increasingly comparable to leading offerings from OpenAI and Anthropic.

From Llama to Muse: a quiet strategic pivot

The Muse family marks a real philosophical shift for Meta’s AI division. The company formalized the transition in April 2026 with the first Muse Spark release, then shipped version 1.1 in July and version 1.2 on August 5, before arriving at 1.3 less than a month later.

The Llama series, which made Meta a favorite among researchers and developers who wanted to run powerful models locally without licensing fees, is no longer the primary vehicle for the company’s frontier ambitions. Muse Spark is closed, API-gated, and paid, placing it structurally in the same lane as GPT-4o and Claude.

That shift matters because it changes how Meta participates in the AI market: instead of relying only on open distribution and downstream developer adoption, it is now pairing model development with a direct commercial access layer. Meta has not abandoned openness entirely. Muse Glimmer, a distilled variant released in August 2026, is designed to run on consumer-grade GPUs and remains open-weight.

$130 billion in bets: what Meta is actually spending

Meta’s capital expenditure guidance for 2026 projects spending between $130 billion and $145 billion on infrastructure this year, directed at data centers, custom silicon, and compute capacity needed to train and serve frontier models.

That scale of spending helps explain why monetization is becoming more explicit. Frontier AI models are expensive to build and operate, and the infrastructure required to support them is now a central part of Meta’s broader product roadmap. Mark Zuckerberg has described the company’s goal as building what he calls “personal superintelligence.” The company already serves roughly three billion people daily across Facebook, Instagram, and WhatsApp, with Meta AI already integrated across its apps as the consumer-facing expression of this strategy.