Meta launches Muse Wonder and Muse Wonder Full, its most capable models yet
On October 12, 2026, Meta unveiled Muse Wonder and Muse Wonder Full, calling them "our most capable models yet" (Meta Newsroom). Muse Wonder is already available through the Meta Model API and Muse Code, while Muse Wonder Full is reserved for Muse Code subscribers and isn't offered via the API (Meta Newsroom; TechCrunch). Both come from the Muse Spark family, launched in April 2026, and were built by Meta Superintelligence Labs (TechCrunch).
The promised capabilities include "Active Memory", the ability to retain relevant information across conversations, and "Active Focus", the ability to stay on goal, try alternative paths and ask the user for clarification (Meta Newsroom). The model was trained at the Prometheus data center in Ohio, operational since 2026 (TechCrunch).
Every performance comparison comes from Meta. According to TechCrunch, the model card reports scores comparable to Fable 5.1 and GPT-6 Astra, plus an edge over Fable, Opus and Astra on MemoryBench, a benchmark for memory and continual learning. TechCrunch gives no figures, and neither does the official post.
According to Euronews, Meta's post cites Artificial Analysis data in which Wonder beats Spark and models from Anthropic, OpenAI and Google, data that also places it among the most expensive models. No independent evaluation has been published so far.
According to TechCrunch, API pricing is $4 per million input tokens and $20 per million output tokens, a cost Meta considers comparable to GPT-6 Sol and Opus 5.5. The price doesn't appear in Meta's official post. For use in the Muse app, monthly subscriptions cost $20 or $100 (CNBC; Euronews, roughly €18 or €90).
Meta hasn't disclosed the parameter count, whether weights will be open, or availability by country or in the EU. The 1-million-token context window appears only in third-party trackers and hasn't been confirmed by Meta.
Meta says the models are trained not to disclose private information without consent and to refuse harmful requests, and that a full safety evaluation report will be published soon (Meta Newsroom). For now the report isn't available, so the safety claims remain unverified.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read Meta's official post on about.fb.com (dates, names, availability, Active Memory/Focus, safety commitments) and two independent outlets: TechCrunch for pricing, MemoryBench and the data center; Euronews/AFP for the Zuckerberg quote, Artificial Analysis and subscriptions. The CNBC page returned a 403 error, so the $20/$100 subscription figure comes from Euronews, which confirms it. OpenRouter listed no prices. Two stories were discarded: the UCLA study (September 29, outside the 7-day window) and the Council of Europe–Microsoft memorandum (non-binding, with little practical content).
- Incertezze
- All performance comparisons come from Meta: no figures in the official post, no independent evaluation. The full safety report hasn't been released yet. Parameter count and availability in Italy/the EU are unknown. The 1-million-token context window appears only in third-party trackers. Euronews calls the larger model 'Wonderful', but its official name is 'Muse Wonder Full'.
- Perché pubblicarla
- It's Meta's first frontier model, from a major player that has been playing catch-up. It marks the definitive shift from Llama's open weights to paid proprietary models and makes 'active memory' a competitive battleground. New ground: our previous Meta story covered Muse Spark and mathematics.