Meta launches Muse Spark 1.1 with first-ever paid API, pricing set to undercut rivals

Meta Launches Muse Spark 1.1 With First-Ever Paid API

Meta released Muse Spark 1.1, its most advanced language model, with a new paid API tier for developers—the company's first monetized AI offering. Mark Zuckerberg pledged to undercut rivals on pricing, positioning the move as a way to compete in a crowded market while capitalizing on Meta's own data center efficiency.

The model shows meaningful improvements in agentic reasoning, tool use, and coding capabilities. Meta employees are already using it internally to build products across the company's app portfolio, which sets up a critical test: how quickly will Meta migrate its own workloads from competitors like Google, Anthropic, and OpenAI onto its own infrastructure.

The Internal Use Case as Validation Signal

Meta's shift matters beyond revenue. If the company can't convince itself to run substantial internal workloads on Muse Spark 1.1, other enterprises will likely hesitate to adopt it. The timing adds pressure—Google recently told Meta it lacks capacity to meet the company's AI demand, a signal that Meta's internal consumption needs are already straining external suppliers.

The economics create a direct incentive for Meta to use its own model. Since Meta owns the data centers and depreciates the hardware regardless, running its own model costs only electricity; using a competitor's API carries margin on top of compute. That spread should push Meta toward aggressive internal adoption and broader token maxing on its own systems.

The Vertical Integration Play

Meta's real advantage sits deeper than API pricing. The company controls both the infrastructure and the end products—particularly its ad network. An image generation model trained on advertiser data becomes a flywheel: ads reveal which creative performs, the model generates new variations based on those signals, and performance data feeds back into model improvement. This feedback loop doesn't exist for general-purpose image generators like ChatGPT or Midjourney, which optimize for user satisfaction rather than advertising performance.

The same logic extends across Meta's ecosystem. Using its own models internally, controlling the training loop, and selling API access creates an economic forcing function: if external customers will pay significantly more for compute than the internal value Meta captures, it signals the model may not be as useful to the company's core business as claimed—or that Meta is underpricing its own services.