Meta's Muse Wonder is its first frontier AI model. What it does and what it costs
Meta has been the big AI lab without a top-tier model of its own. That changed on Monday. Its Superintelligence Lab released Muse Wonder and a larger Muse Wonder Full, which TechCrunch calls Meta's first frontier model release. Here is what is new, what it costs and what has not been checked yet.
Disclosure: this site's posts are drafted with AI help. This one was drafted with Claude, made by Anthropic, a Meta competitor. It sticks to published reporting and Meta's own statements.
What Meta released
- Two models. Muse Wonder and Muse Wonder Full are bigger, more capable successors to Muse Spark, Meta's earlier lightweight, fast model, TechCrunch reports.
- Memory is the headline. Meta calls it Active Memory: the model keeps important context about you across conversations and pulls up what matters for the task at hand. A Meta research scientist said in the launch video that the goal is for it to remember what matters to you "over days, weeks, and months."
- Sticking with a goal. A second feature, Active Focus, is meant for long jobs. Meta says the model tries other approaches when stuck, adapts to new information, asks when it needs clarification and checks in before important actions.
- Where it was trained. Meta's Prometheus data center in central Ohio, which came online this year, according to TechCrunch.
- Muse Wonder: The one developers can use today. Where: Meta Model API and Muse Code now; the Muse app soon, subscribers first. API price: $4 per million input tokens, $20 per million output tokens. Built for: Personal and work agents, memory, coding.
- Muse Wonder Full: The bigger version. Where: Muse Code subscribers only. API price: Not offered through the API. Built for: The hardest long-running tasks, per Meta.
Sources: TechCrunch, Oct 12, 2026; Meta AI Research. Performance claims are Meta's own.
How good is it?
Meta says the Wonder models score in the same range as the newest top models from Anthropic and OpenAI, and that Muse Wonder beats them on MemoryBench, a test of how well a model builds and uses memory over time, TechCrunch reports. Meta's research post also cites strong results on a legal-agent benchmark and long coding tasks.
Two cautions. TechCrunch did not give exact scores, and these are Meta's own numbers. Independent testing usually takes a few days to a few weeks, so treat "state of the art" as a claim for now.
What it costs
Through Meta's API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. TechCrunch puts that roughly in line with the top models from OpenAI and Anthropic. In plain terms: a million tokens is several long books, so light personal use costs cents; it adds up only for apps and heavy agent work.
Example cost
- Short chat: $0.03 (company stated)
- Summarize a long report: $0.24 (company stated)
- Long agent task: $2.00 (company stated)
Our arithmetic from the API prices reported by TechCrunch, Oct 12, 2026. Token counts are rough examples.
Why it matters
Memory is where personal AI assistants are heading. An assistant that remembers your projects, preferences and people is more useful, and it also knows a lot more about you. Meta's business runs on ads, so it is worth reading Muse's privacy and memory settings before telling it everything.
What to watch
- Independent benchmark results over the next couple of weeks.
- When Muse Wonder reaches the Muse app, and what Muse subscriptions cost.
Sources
- Meta releases its first frontier model, Muse Wonder, TechCrunch, Oct 12, 2026
- Muse Wonder, Meta AI Research, Oct 12, 2026