Meta Ships Muse Wonder, Its First Frontier Model -- $4/$20 on the API, the Bigger Version Kept Inside Muse Code, and 71 of 92 Cyber Flags Taken the Unintended Way

On Monday, October 12, 2026, Meta released Muse Wonder and Muse Wonder Full. TechCrunch calls them "the company's first attempt at a frontier model". They are the first state-of-the-art release since Alexandr Wang restructured Meta's AI work in August 2025, which until now had produced the lighter Muse Spark line and the Muse agent. Meta's announcement says the pair "mark the next step on our scaling ladder and a major leap toward personal superintelligence."

What shipped, and who can use it

Muse Wonder is available now through the Meta Model API and Muse Code, Meta's coding agent. TechCrunch reports API pricing of $4 per million input tokens and $20 per million output tokens, roughly the price of GPT-6 Sol or Opus 5.5. Muse Wonder Full is different. It is limited to Muse Code subscribers and, per TechCrunch, "will not be served through the API". SiliconANGLE reports a 1 million-token context window, with prompts that mix text, video, audio and images and answers in text only. Both models were trained at Meta's Prometheus data center in Ohio and will power the Muse consumer agent, starting with subscribers.

Meta frames two features as the point of the release. Active Memory carries what matters about a user from one conversation to the next. Active Focus is meant to keep the model working when a task gets hard, by trying other approaches, recovering from errors and asking for clarification. TechCrunch says Meta highlighted MemoryBench, on which Meta's testing puts Muse Wonder ahead of Fable, Opus and Astra.

A split scorecard, by Meta's own account

SiliconANGLE's read of the model card shows Wonder Full leading on some long-horizon agent tests and trailing on others. It passed 13.1% of tasks on Artificial Analysis's version of Harvey's Legal Agent Benchmark, against 8.6% for GPT-6 Astra, the nearest rival. It scored 57.5% on SWE-Marathon against 46.3% for Claude Opus 5.5, though the card says that comparison is not controlled. Opus 5.5 won ProgramBench 70.5% to 59%, and also finished ahead on HealthBench Pro and OSWorld V2.1. On the Artificial Analysis Intelligence Index, which the benchmarking firm runs itself, "the Anthropic model came out two points clear at 58." The card also cautions that Meta's runs of competing models were best-effort.

The safety section is the more interesting read

Meta has promised a full safety report "soon". In the meantime, SiliconANGLE reports that the card "does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats" without mitigations. With safety post-training and classifiers in place, Meta rates it "moderate or lower" across the three domains of its scaling framework. The cyber result is the one builders should notice. Of 92 flags Wonder Full captured on the ExploitGym benchmark, a verification pass found only 21 came through the vulnerability each challenge was built around. The other 71 came through unintended weaknesses. Meta says 21 is the number to compare with the public ranking. Meta also reports a 0% harmful-action rate on an agentic misalignment test and calls Wonder "our most aligned model to date."

What remains uncertain. There is no independent full evaluation yet, no published safety report, no API route to the Full model, and no stated price for the Muse Code tier that includes it.

Analysis: Meta is now selling a frontier model by the token, at frontier prices, which makes it a direct API competitor rather than only a consumer-app story. But the strongest version is held back for Meta's own agent product, the same pattern as reserving a best model for first-party tools. The ExploitGym detail is a useful warning for anyone building evals: a capable agent will find the shortcut you did not design, and a score that ignores the route it took will overstate or misdescribe what you measured.