Meta's Best New Model Is Closed, Paid, and US-Only

What Meta actually launched

Meta Superintelligence Labs released Muse Spark 1.1 on Thursday, 9 July 2026, a multimodal reasoning model built for agentic and coding work, with a self-managed one-million-token context window and native support for orchestrating sub-agents, MCP tools, and custom skills. Alongside it Meta opened the Meta Model API, the company's first paid, hosted, pay-as-you-go endpoint, priced at 1.25 dollars per million input tokens and 4.25 dollars per million output tokens, with 20 dollars of free credit for a new account. Consumers can use the model free inside the Meta AI app in its Thinking mode; developers pay per token.

The detail that matters is what you cannot do with it. Muse Spark 1.1 is closed-weight and proprietary: there is no download, no local deployment, no fine-tuning on your own hardware. That is a clean break from the Llama family, whose open weights let any team run the model on its own servers with no vendor in the loop. For the first time, Meta's best model is something you access, not something you hold.

Meta did not abandon open weights. It split the ladder.

Read the move as a two-tier strategy, not a reversal. Llama stays open and general-purpose; Muse Spark becomes the closed, premium, agentic tier. Meta looked at the hosted-inference market that OpenAI and Anthropic built and decided it was too large to cede, so it kept the open weights as the commodity floor and put its frontier work behind a meter. The company that made open weights a strategy now runs the same paid-API business it once positioned itself against, on the models it cares most about.

That reframes what an open-weights release from a commercial vendor actually promises. It promises a good-enough, self-hostable floor, not first claim on the frontier. The best model and the free model have started to diverge, and the gap is exactly the agentic, long-context, tool-using capability that owners most want to build on. An open license on last generation's weights is not the same as open access to this generation's best.

Why the US-only preview is the sharper problem for Europe

Availability, not price, is the immediate constraint here. The Meta Model API preview is limited to US developers, so a team in Munich, Milan, or Manchester cannot yet call the new flagship at all, and when access widens the only supported route will be a US-hosted, metered endpoint. A European team that adopted Llama precisely to keep models and data on its own infrastructure gains nothing from Muse Spark's frontier features without sending its traffic to a US API. The independence that open weights bought does not extend to the tier now sitting behind the paywall.

This is the quiet cost of leaning on any one vendor's free posture. Free-tier and open-weights strategies from commercial companies are snapshots of a business decision, and business decisions revert to monetization when the market is large enough. Betting a stack on one supplier staying generous is a base-rate error; the base rate is that the best capability migrates to the paid product.

What to actually do about it

The protection is architectural, and it is cheap to build now. Put a thin abstraction layer between your application and whatever model serves it, so the model is a configuration value rather than a hard-coded dependency. Keep a self-hostable open-weights model, Llama or otherwise, as your floor for the work that does not need the frontier, and route only the tasks that genuinely need agentic reasoning to a paid API. Then a vendor moving its best tier behind a meter, or restricting it by region, is a routing change and a cost line, not a re-platforming project.

The owners who will feel this move least are the ones who never assumed any single model was permanent. They treat every model, open or closed, as swappable, price the frontier tier as an operating cost rather than a free input, and keep an open-weights option warm so they always have a floor they control. Muse Spark 1.1 is not a warning that open weights are ending; it is a reminder that a vendor's best work and its free work are two different things, and only one of them is yours to keep.

Read next: Alibaba Bans Claude Code as a Backdoor Risk  |  Meta's Image AI Now Searches the Web to Draw