Meta ships Muse Spark 1.1 and starts charging for its own model

Meta released Muse Spark 1.1, describing it as its most capable model yet for real-world coding and agentic tasks. That superlative is Meta's own framing on a vendor benchmark, so treat it as a claim, not an independent verdict. The bigger news sits alongside the model itself: Meta stood up the new Meta Model API and, for the first time, is billing developers to run its own model. Exact pricing was not confirmed in early reporting, so there are no numbers here yet.

In short, two things shipped at once: a new flagship model aimed squarely at agentic and coding workloads, and a paid pipe to reach it. The second is the story. Meta has released capable models before; it has never charged for its own model before.

Meta built its identity on open weights, free to run, in the Llama era. A paid, hosted API is a move towards monetising inference directly, the same business OpenAI and Anthropic run. Read it as a hedge: as open-weight rivals like DeepSeek, Meituan LongCat and Qwen commoditize the openness Meta pioneered, giving models away stops looking like a business on its own. Whether Meta keeps shipping open weights alongside the paid API, or quietly retires the free-to-run track, is the open question buyers should watch. For now the direction is clear enough: the open-source champion has put its newest model behind a meter.

The timing matters. Muse Spark 1.1 lands right after Meta's AI-focused restructuring, which cut about 8,000 roles and reassigned about 7,000 to AI teams, making it an early proof-point that the painful reorg is meant to pay off. Treat that link as context, not a claimed causal win yet. It also drops into the busiest launch fortnight of the year, sharpening the question of whether Meta can differentiate on agentic and coding reliability against fresh rivals: