Meta releases Muse Wonder and Wonder Full, its first frontier models

Meta released two models on October 12, 2026: Muse Wonder and Muse Wonder Full, which it calls its most capable models to date and 'frontier models for agents'. Muse Wonder is available now to developers through the Meta Model API and through Muse Code; Muse Wonder Full is available now only to Muse Code subscribers. Meta says Wonder will later power the consumer Muse app, rolling out to paying subscribers first.

What it costs

TechCrunch reports the Meta Model API price for Muse Wonder as $4 per million input tokens and $20 per million output tokens, and describes that as roughly the cost of GPT-6 Sol or Claude Opus 5.5. Meta's announcement does not state a price. Wonder Full is not served through the API at all, according to TechCrunch: it is a Muse Code subscription feature. For a team, that is the practical split. Wonder is a drop-in candidate for an API routing test at a familiar rate; Full is a per-seat coding product, and there is no way to put it behind your own gateway.

What Meta claims, and who measured it

Meta's announcement says the models reach 'state-of-the-art performance on personal and professional agentic tasks', and names two capabilities: Active Memory, which carries details across conversations, and Active Focus, which Meta describes as continuing toward a goal when a task gets hard, trying other approaches, recovering from errors, and checking in before important actions. Meta's research post names three benchmarks it ran, FocusBench, SWE-Marathon and Harvey's Legal Agent Benchmark, and says the full results table is in its report; the post itself gives no scores.

TechCrunch, citing Meta's model card, reports that Wonder and Wonder Full score comparably to Fable 5.1 and GPT-6 Astra, and that Wonder scores higher than Fable, Opus and Astra on MemoryBench, a memory and continual-learning benchmark. All of those numbers are Meta's own testing. Nobody outside Meta has published an evaluation yet, and Meta's safety report with additional evaluations is described as 'available soon'.

What is not in the announcement

• No context window length.
• No list of supported modalities (text, image, audio, video).
• No statement on open weights; nothing suggests these are open models.
• No named customers, and no benchmark scores on the public pages.

Why it matters

TechCrunch calls this Meta's first frontier release since Alexandr Wang restructured the company's AI effort in August 2025; the earlier Muse Spark, announced on April 8, 2026, was a lower-latency model. Meta says the Wonder models were trained, evaluated and released under its Meta Superintelligence Scaling Framework, with what it calls its strongest safeguards yet against chemical, biological, radiological, nuclear and cyber misuse, while keeping defensive security capabilities for teams that want to use the models to find flaws in their own code.

What to do

If you already run a bake-off across GPT-6 Sol, Claude Opus 5.5 and Fable 5.1, add Muse Wonder through the Meta Model API at the published rate and score it on your own tasks before reading Meta's table. Do not budget for Wonder Full in an API-based system; it is subscriber-only. Hold procurement decisions that depend on safety evaluations until Meta publishes the full report it has promised.