For most of the artificial intelligence boom, Meta’s model strategy had the shape of a giveaway. The Llama family of open models was free to download and free to modify, and the company argued that the generosity was the point: an open ecosystem would carry Meta’s technology further than any paid product could. On Wednesday the company put its newest flagship on a different road.
Meta released Muse Spark 1.3, an update to the model family aimed at the two uses that have come to define the frontier: writing software and running autonomous agents. The company says the new version handles longer tasks, juggles several jobs at once and follows complex instructions more reliably than its predecessor. Compared with Muse Spark 1.2, it makes about 20 percent fewer tool calls and consumes about 25 percent fewer tokens on coding tasks, and Meta says it strengthened the model’s defenses against adversarial inputs and prompt injection, the attacks that try to make a system ignore its own instructions. The model is available through Muse Code and the Meta Model API, with the highest-reasoning mode to follow once safety testing is complete. The release also signals where Meta believes the next battles will be fought: Muse Code is the company’s answer to a generation of tools that put a model beside the keyboard, and agentic workloads are where the largest cloud providers now say the growth is.
The technical claims were not what drew the most attention. For the first time, Meta is charging for access to its flagship model instead of giving it away: the top tiers of Muse Spark on the company’s API carry a price. The company said it still expects to release open weights of Muse Spark and larger models in the future, a concession that keeps part of the old promise alive even as the paid product takes the leading role. But the direction is clear, and Bloomberg’s analysis framed it as a change in Meta’s business model: after years of trading open models for ecosystem influence, the company is moving toward selling its best technology directly.
The economics behind the shift are easy to state. Serving the largest models to millions of developers costs real money, because every request consumes expensive computing, and the frontier of model quality has moved to sizes that only the best-funded companies can train and operate. OpenAI and Anthropic have charged for their strongest models from the start, and every major lab has moved in the same direction as costs have climbed. Meta was the exception, the one company that gave its best work away, and the exception has now ended for its newest line.
What makes the decision delicate is the goodwill that the free strategy accumulated. For years, developers built businesses, tools and careers on Meta’s open models, treating Llama as the dependable free option in a field of paid services, and the company leaned on that trust when it argued that openness would win it the developer mindshare that closed ecosystems could not buy. The reaction from the open-source community to Wednesday’s announcement showed how much of that goodwill was banked, and how quickly it can be spent when a flagship model appears behind a paywall.
There is a distinction that may soften the argument, at least for some. Llama itself remains open, and the paid tier applies to the newer Muse Spark line; the company is not closing off the models that built its reputation. Whether the community reads the release as a one-time exception or as the shape of things to come will depend on what Meta does next, and the promised open weights of Muse Spark will be watched as the first test of whether the company means to straddle both worlds or to move gradually from one to the other.
The bet carries risk in both directions. If developers pay for the capability, convenience and support that a hosted flagship provides, Meta gains a revenue stream to set against the enormous cost of frontier AI, and the free models keep feeding the ecosystem that feeds it in return. If they do not, and rivals offering open models undercut the paid product, Meta could find itself paying a different price: the gradual erosion of the default position it built by giving things away.
The stakes go beyond one model release. Meta has spent years arguing that open development is the future of AI, that the technology is too important to be controlled by a few companies, and that its own openness would keep it close to the developers who decide which platforms win. Charging for the flagship does not contradict that argument by itself, but it blurs the line between the company’s stated philosophy and the commercial reality it now faces.
Muse Spark 1.3 is faster, cheaper to run and, for the first time, not free. Meta spent years betting that openness would build its ecosystem; with this release it is testing how much of that bet it can cash in without losing the trust that made the bet work in the first place.
