Meta Muse Wonder: a closed frontier model priced at $4 per million input tokens

On October 12, Meta introduced two new large language models, Muse Wonder and Muse Wonder Full, built by its lab Meta Superintelligence Labs. The company presents them as its most capable models for personal and professional "agents," the assistants able to chain several tasks on their own. Unlike the older Llama family, these models are not open-weight: they cannot be downloaded or installed in-house. They are only available through Meta's paid services.

What Meta highlights

On its research blog, Meta stresses two features. The first, Active Memory, keeps the user's context from one conversation to the next. The second, Active Focus, is meant to push the model to persevere toward a goal, try other approaches, ask for clarification and check before any important action. Meta also claims strong results in law, finance, health and long-running development projects. The models were trained in its Prometheus data center.

These results come from Meta's own measurements, including two internally designed tests. Next considers that the highlighted features are partly marketing, since memory already exists at competitors. On safety, Meta says it has its strongest protections to date against misuse in the CBRN (chemical, biological, radiological, nuclear) and cyber domains. It also says it has kept capabilities useful for cyberdefense. The full safety report is announced as "coming soon": it had not been published as of October 13.

What an independent evaluation shows

The firm Artificial Analysis measured both models on its Intelligence Index (version 4.3, maximum reasoning effort). Muse Wonder Full scores 56 and Muse Wonder 52, compared with 48 for Meta's previous model, Muse Spark 1.3. On the same index, Anthropic's Claude Opus 5.5 reaches 58, according to Next.

The evaluation also points out weaknesses. On the AA-Omniscience knowledge test, Wonder Full scores 30, far from Claude Opus 5.5's 46. It does improve by 5 points over Muse Spark 1.3, thanks to a hallucination rate cut from 33% to 23%. Another limitation: Wonder Full is only available to Muse Code subscribers, which makes it hard for third parties to evaluate.

Pricing and availability in Europe

• API: only Muse Wonder is offered. The price is $4 per million input tokens and $20 per million output tokens, with a context window of one million tokens. This is the same rate as Claude Opus 5.5, according to Artificial Analysis.
• Muse Code: Wonder Full is reserved for the "Heavy use" plan, priced at €42.99 per month in France according to Next.
• Muse app: Meta says Wonder is coming "soon," first for subscribers. According to several media outlets, the app is not currently offered in Europe.
• Open weights: only Muse Glimmer, a more modest distilled model with 30 billion parameters, is released as open weight.

Key takeaways

For a European company, Muse Wonder is one more option among closed models available through an API, not an alternative to self-host. As The Next Web points out, Meta has not signed the code of practice provided under the EU AI Act for general-purpose AI models. Before any use, compare the model on your own use cases, not on the vendor's charts. Also read the data processing and data location terms closely.