On 12 October, Meta released two new models: Muse Wonder and Muse Wonder Full. They are the strongest models Meta has built so far. That makes Meta worth seriously considering alongside OpenAI, Anthropic and Google for demanding agent work. Below you can read what they cost, how good they are, and what to watch out for before you put work data into them.
What has been released?
According to Meta's announcement, the models are built for agent tasks, such as handling an insurance claim, reading through a rental contract, producing financial analyses and writing code. Meta highlights two new features. With Active Memory, the model remembers what matters to you, including across conversations. With Active Focus, the model keeps working towards your goal even when a task gets difficult.
The two variants are available in different ways:
- Muse Wonder can be used via the Meta Model API (model ID `muse-wonder-1.0`) and in Muse Code. You pay $4 per million input tokens and $20 per million output tokens. Cached input costs $0.20 per million tokens. The context window is 1 million tokens.
- Muse Wonder Full is not available through the API. You can only use it with a Muse Code subscription. According to VentureBeat, that costs $50 per month (Power Usage) or $200 per month (Maximum Usage).
Both models are closed. Unlike Llama, you therefore cannot download them or host them yourself.
How good are they?
Meta mainly publishes figures from its own and specialised benchmarks. VentureBeat set out the independent Intelligence Index from Artificial Analysis. Muse Wonder Full scores 56 there, the same as Claude Sonnet 5.5 and above GPT-6 Astra (53). Claude Opus 5.5 remains highest at 58. The standard Muse Wonder scores 52. All scores were measured with maximum reasoning effort. On a coding test for long tasks (SWE-Marathon) and a legal test (Harvey LAB-AA), Wonder Full scores highest according to VentureBeat.
Also look at the price. In our OpenRouter data, Claude Opus 5.5 costs exactly the same: $4 input and $20 output per million tokens. Meta's model that you can use via the API is therefore not cheaper than Opus 5.5, while it scores 6 points lower on that index. Compare prices and scores at /modellen/.
Privacy: check the default setting
According to VentureBeat, Meta does not use what you send via the API to train models by default. In Muse Code, by contrast, "contributor mode" is switched on by default. That allows Meta to use your prompts, code, files and outputs to improve future models. Zero data retention is only available to API organisations that Meta approves for it.
Meta's announcement says nothing about storage in the EU or availability in Europe. Meta has also not yet published its full safety report. If you work with client or patient data, as with lawyers, insurers or in healthcare, wait until there is a data processing agreement and clarity about where data is stored. You can see how other models handle this at /veiligheid/.
For software developers who want to try Muse Code: switch off contributor mode first, and do this before you share company code.
In brief
- Muse Wonder costs $4/$20 per million tokens via the API, the same as Claude Opus 5.5. You only use Muse Wonder Full through Muse Code ($50 or $200 per month).
- In the Artificial Analysis index, Wonder Full scores 56 and the API version scores 52. Claude Opus 5.5 reaches 58.
- In Muse Code, training on your input is switched on by default. Meta has said nothing yet about storage in the EU.
