Meta's Muse Wonder Is Here, and It Costs Exactly What Opus Costs
Meta's first frontier models are built for agents that remember you and keep working the problem. The Neuron's October 13 edition breaks down the scores, the Opus level price, and the training setting hiding in Muse Code.
Built for agents, not one shot answers
Meta released Muse Wonder and Muse Wonder Full, the company's first frontier models, built to do multi step work for you instead of answering once and stopping. Think planning a schedule, working through a claim, or juggling the rules of a wedding seating chart until every guest fits.
The pitch rests on two features. Active Memory keeps context about you across conversations, so you stop explaining your budget, your lease, or your project at the start of every new chat. Active Focus keeps an agent pushing toward the goal when it hits a wall, tries another route, and asks you before anything consequential happens.
Second place at first place prices
On Meta's own scorecard, Wonder Full finishes ahead of Claude Opus 5.5, Fable 5.1, and GPT 6 Astra on 7 of 14 rows, mostly long software projects and legal agent tasks. On the independent Artificial Analysis index, the order flips: Opus 5.5 scores 58, Wonder Full 56, GPT 6 Astra 53, and the standard Wonder 52.
The price is the story. Wonder costs $4 per million input tokens and $20 per million output tokens, the same as Opus 5.5, the model that outscores it. Sonnet 5.5 costs half as much and also scores higher. One launch day test made the point in a funnier way: a developer set Wonder loose on Pokemon Red, and after roughly 792 turns and about $96, the agent was still wandering Viridian Forest without a gym badge. That is one run, not a benchmark, but persistence without progress is exactly what Active Focus promises to prevent.
Where you can actually get it
Access is split, and the split matters if you are budgeting for production work:
- Muse Wonder is live on Meta's Model API and inside Muse Code, Meta's terminal app
- Muse Wonder Full is only sold inside Muse Code plans at $50 and $200 per month, never by the token
- The Muse app is coming later, subscribers first
- Neither model ships downloadable weights, and the open weights Meta promised for Muse Spark still have not arrived
Check one setting before you paste client work
A model that remembers you is only as safe as what the company does with that memory. VentureBeat reports that Muse Code defaults to a Contributor mode that lets Meta train on what you submit. You can switch it off, but the change only covers future submissions. The API does not train on your content.
For creators and studios, the takeaway is practical. Agent memory could end the daily ritual of briefing your tools on every project from zero, and an agent that keeps working a hard problem is worth real money on production tasks. Just route client documents through the API, read the training setting before your first session, and grade Meta's chart against the independent index before you move any budget.
Sources and further reading
Sources were checked for this article. Tool features, prices and availability can change. Product claims are attributed to their publishers.
