Meta Superintelligence Labs pushed Muse Spark 1.1 on Thursday. The feature list is what you would expect from a Q3 2026 frontier release: multimodal reasoning, million-token context, actively managed context compaction, computer use, multi-agent orchestration where the main agent plans and delegates to subagents that know when to escalate back. There is a coding-agent demo called OpenCode that looks like every other coding-agent demo. There is a benchmark chart that beats some things and loses to others. Fine.
The interesting number is the meter. $1.25 per million input tokens. $4.25 per million output tokens. Public preview of a paid Meta Model API. US developers in immediately, waitlist for the rest. That pricing is aggressively below GPT-5.6 Terra, comfortably below Opus 4.8, and structurally above the actually-free DeepSeek and Qwen tokens the enterprise routing layer has been shifting toward all year. Meta is not undercutting China. Meta is undercutting America.
The context that makes this weird is Meta’s entire prior AI identity. Llama 2, 3, 4, all the Medium-and-Small-and-Whatever variants, the whole four-year argument that open weights were the correct default and that the moat was distribution not model access. That argument was, whatever else you thought of it, a real argument. It is now, functionally, dead. Muse Spark 1.1 does not have open weights. It has a pricing page. Meta trained a closed frontier model, hid it behind an API, put a meter on the API, and shipped it as the flagship of a rebranded org.
Alexandr Wang is running the org. That is the other thing that changed. Wang came in from Scale AI at the top of the Meta Superintelligence Labs restructuring, and he is running a different playbook than the one Yann LeCun spent half a decade defending in public. This is not a value-neutral change of direction. The bet is that closed frontier plus aggressive pricing plus computer use plus multi-agent orchestration plus a million-token context is a product, and the free-weights posture is a giveaway. Meta wants revenue. Meta wants developer lock-in. Meta wants the model to be the reason you built on Meta, not a byproduct of the reason you built on Meta.
Whether it works depends on whether $1.25 / $4.25 tokens with agentic bells is enough to pull anyone away from the two things they were probably already doing. Which are: paying full freight for Claude or GPT-5.6 on the high-stakes tasks, and paying near-zero for a Chinese model on everything else. Muse Spark is neither of those. It is a middle price for a mid-to-frontier capability from a vendor that everyone knows might change its mind about distribution again in six months, depending on which Meta division wins the next org fight.
Nobody who built on Llama for four years is looking at the pricing page and not thinking about that.