Meta Puts a Price Tag on AI for the First Time With Muse Spark 1.1

Meta has never charged for access to one of its AI models before. That changes with Muse Spark 1.1, a new release from Meta Superintelligence Labs that ships with a public API and a per-token price: $1.25 for every million input tokens and $4.25 for every million output tokens. Mark Zuckerberg announced the launch on X, marking a notable departure from Meta's long-standing open-weights posture with the Llama family.

Muse Spark 1.1 is the second model to come out of Meta Superintelligence Labs, following the earlier Muse release, and it arrives alongside Muse Image, the lab's first foray into image generation. Unlike Llama, Spark 1.1 keeps its weights closed and exposes an OpenAI-compatible endpoint, a design choice that lowers the switching cost for developers already building on top of GPT-style APIs. That compatibility signals Meta is competing less on ideology and more on practical developer adoption, meeting engineers where their existing tooling already lives.

A Commercial Pivot Wrapped in Infrastructure Strategy

The pricing move doesn't exist in isolation. Meta is simultaneously building out Meta Compute, an internal cloud unit designed to resell surplus AI infrastructure capacity to outside customers. Pair that with the company's custom MTIA silicon, which is ramping toward production, and a clearer picture emerges: Meta is positioning itself to become an infrastructure and inference vendor, not just a model publisher giving away weights to seed an ecosystem.

Charging for Spark 1.1 gives Meta a live commercial testbed for that ambition. Selling API access at scale generates real usage data on latency, margins, and demand elasticity, all of which feed directly into how Meta prices and provisions Meta Compute capacity down the line. It also means the custom MTIA chips need production workloads to justify the capital expenditure, and a monetized model line supplies exactly that kind of steady, billable traffic.

Why the Shift Matters

For years, Meta's argument against charging for AI access was strategic: open models drove developer mindshare and commoditized the layer above, where rivals like OpenAI and Anthropic were trying to build margin. Muse Spark 1.1 suggests that calculus has shifted. With Meta Compute and MTIA both maturing, Meta appears to be hedging its bets, keeping Llama's open ecosystem play alive while testing whether a closed, metered, developer-friendly product can compete directly with commercial API providers on price and performance.

The OpenAI-compatible endpoint is the detail worth watching most closely. It's a low-friction invitation for developers to test Spark 1.1 against incumbent models without rewriting their integration layer, and it turns Meta's entrance into paid AI services into an easy A/B test rather than a migration decision. Whether that translates into meaningful API market share depends on benchmarks and reliability data that have yet to surface publicly.