What was announced
Models: Muse Wonder and Muse Wonder Full, from Meta Superintelligence Labs. Code-named Watermelon.
Price: $4.00 input / $20.00 output per million tokens, Meta Model API, Wonder only.
Wonder Full: Muse Code subscribers only. Not available through the API.
Not published: parameter count, context window, numeric benchmark scores, open-weights status.
The price is the only hard number in the announcement
Muse Wonder costs $4.00 per million input tokens and $20.00 per million output through the Meta Model API. That is the single verifiable figure Meta put in front of buyers on launch day.
Here is where it sits against the models Meta says it matches.
| Model |
Input / 1M |
Output / 1M |
| Meta Muse Wonder | $4.00 | $20.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Mistral Large 4 (list) | $1.36 | $4.18 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
Twice GPT-6.1 Sol. Twice Claude Sonnet 5.5. Forty times Claude Haiku 5.5 on output.
That is a confident price from a company that spent the last two years being described as behind. It is defensible if the model is genuinely better — and that is precisely the thing nobody outside Meta can yet check.
What Meta claims
Alexandr Wang, who heads Meta Superintelligence Labs, said the models are "true frontier models, competitive with the best models in the world."
Meta's model card states that Wonder and Wonder Full score comparably to Fable 5.1 and GPT-6 Astra, and that on MemoryBench — which tests continual learning and memory construction — Wonder scores higher than Fable, Opus and Astra.
Every one of those is a comparison without a number. "Comparably to" and "higher than" are rankings, not measurements. No score, no margin, no test configuration. A model card that places you at the frontier without saying by how much is a claim, not evidence — and at double the incumbent price, the burden of proof sits with the vendor.
The two features that are actually new
Active Memory. The model retains information a user has given it across conversations and retrieves it when relevant. Research scientist Kavya Kopparapu framed it in the launch video as making sure the model remembers what matters to you.
Active Focus. Keeps an agent working toward a goal, trying different approaches and adapting when it hits an obstacle rather than stopping or drifting.
Both target the same weakness: agents that lose the thread. If Active Memory works as described, it is the more consequential of the two — persistent context across sessions is the gap between a chatbot and an assistant, and MemoryBench is the one benchmark where Meta claims a lead rather than parity.
What Meta did not say
- Parameter count. Not published for either model.
- Context window. Not published.
- Any numeric benchmark score. Comparisons only.
- Whether Wonder will ever be open-weight. Wang said "we're still committed to open source" and pointed at Muse Glimmer, which is open-weight. He did not extend that to Wonder.
Meta published a model card on Monday and says a fuller report follows. Until that lands, every performance claim here rests on Meta's own testing.
Where it came from
The models were trained at Meta's Prometheus data centre in central Ohio, which came online earlier this year. Meta says it strengthened safeguards against cyber and biological misuse and added protections for personal information — the latter being a necessary consequence of Active Memory, since a model that remembers your details is a model that stores them.
Should you use it
Running agents that need memory across sessions? This is the one case where Wonder has a specific claimed advantage. Worth a trial against your own workload.
Costing out general API usage? Not yet. You are paying double for parity claims you cannot verify. Sol and Sonnet do the same job at half the price today.
Already on Muse Code? Wonder Full is included and not sold separately, so there is nothing to decide — use it and see.
Optimising for cost? Not remotely. Claude Haiku 5.5 is $0.50 per million output against Wonder's $20. If the task does not need a frontier model, this is the wrong forty-fold.
What to watch
Independent benchmarks. Meta has made a parity claim at a premium price, and the next two weeks of third-party testing will either support it or not. That is the whole story, and nothing before then settles it.
Worth noting in Meta's favour: publishing a model card on launch day and promising a fuller report is more than some vendors do. The complaint here is the absence of numbers, not the absence of a document.
FAQ
How much does Meta Muse Wonder cost?
$4.00 per million input tokens and $20.00 per million output through the Meta Model API. Muse Wonder Full is not sold through the API — it is included with a Muse Code subscription.
Is Muse Wonder open source?
Not announced. Alexandr Wang said Meta remains committed to open source and pointed to Muse Glimmer, an open-weight model, but made no commitment about Wonder.
How does it compare to GPT-6.1 Sol and Claude Sonnet 5.5?
Meta's model card claims comparable performance to Fable 5.1 and GPT-6 Astra. It publishes no numeric scores. On price, Wonder is exactly double both Sol and Sonnet, which run at $2 input and $10 output.
What is MemoryBench?
A benchmark testing continual learning and memory construction. It is the one test where Meta claims Wonder beats rather than matches its rivals, though again without publishing a score.
What is the difference between Muse Wonder and Muse Wonder Full?
Wonder is available through the Meta Model API and Muse Code. Wonder Full is restricted to Muse Code subscribers and is not served through the API.
How many parameters does it have?
Meta has not published a parameter count or a context window for either model.