Meta releases Muse Wonder, its first frontier model for agents
Meta has released Muse Wonder and Muse Wonder Full, the most capable models its Superintelligence Labs has shipped. They're built for agents that do long, real tasks: reviewing a lease, handling an insurance claim, building an app. Wonder is on Meta's API now at $4 and $20 per million tokens; the stronger Full is only in Meta's paid coding tool, and neither comes with open weights.
What happened
On October 12, Meta introduced Muse Wonder and Muse Wonder Full, built on its earlier Muse Spark model and aimed at "personal and professional agentic tasks". Meta's examples are the kind of job that takes many steps: handling an insurance claim, reviewing a lease, building an app. It names law, finance, health and coding as the target fields.
Two features get names. Active Memory carries relevant context from one conversation to the next. Active Focus is about staying on a hard task: working past errors, trying another route and checking in before an important action.
Where you can use it
- Muse Wonder: on the Meta Model API as
muse-wonder-1.0, labelled a preview, and in Muse Code, Meta's coding agent. - Muse Wonder Full: only in Muse Code, on the two highest usage plans. It isn't on the API.
- The Muse app: Meta says Wonder will "soon" power more of Muse, its personal assistant app, starting with subscribers. No date.
According to the model card, it takes text, images, video, audio and PDFs, answers in text, and has a 1 million-token context window. Meta hasn't released the weights; its open-weight model is a separate one, Muse Glimmer.
What it costs
On the Meta Model API, Muse Wonder costs $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens, with no surcharge for long context. Web search through the API adds $2.50 per 1,000 searches. That's the same list price Google has announced for Gemini 4 Argon after its introductory period.
Muse Code has four usage tiers, from Everyday to Maximum, and Meta's documentation doesn't list their prices. We'll add them once Meta publishes them.
How good is it
Meta's model card compares Wonder and Full with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, and claims state-of-the-art results on agentic work, with the clearest leads on legal and long coding tasks. It doesn't include Gemini in the comparison. On the independent Artificial Analysis index, as reported by VentureBeat, Wonder Full scores just below Opus 5.5 and above Gemini 4 Argon, while the standard Wonder, the one on the API, lands below Argon.
Meta says a full safety evaluation report will follow.
What it means for your prompts
- Give it the whole job. It's built for multi-step tasks, so describe the outcome, the constraints and when it should stop and ask, not just the first step.
- Attach the documents. With PDFs, images and a 1M-token context, paste the lease or the claim itself instead of summarizing it.
- Ask it to check in. Active Focus is meant to pause before important actions; say which actions count as important for you.
Try it
I've attached [document type, e.g. my apartment lease]. I'm the [your side, e.g. tenant] in [country or state]. Work through it in this order: 1) a one-paragraph summary of what I'm agreeing to, 2) every clause that is unusual, one-sided or costly for me, with the clause number, why it matters and what I could ask to change, 3) deadlines and notice periods as a dated list, 4) anything that seems to be missing. Quote the clause text you rely on. Before giving any legal conclusion, tell me what you'd need to know about local law to be sure, and stop to ask me if a clause can be read two ways.
Sources
- Meta, Introducing Muse Wonder and Muse Wonder Full, October 12, 2026: models, purpose, availability
- Meta, Muse Wonder research post and model card: inputs, context window, comparisons
- Meta, Model API models and pricing: model ID, preview status, prices
- VentureBeat, Meta's new Muse Wonder models, October 12, 2026: Artificial Analysis scores