Meta announced two new AI models on 12 October 2026, Muse Wonder and Muse Wonder Full, describing them as its most capable models to date, with a particular focus on agents that carry memory of a task or a person across separate conversations rather than starting fresh every time. For a small or medium business already using AI tools for admin, customer queries or coding help, the model itself matters less than one specific capability: whether it can actually remember your business's context from one session to the next.
What happened
Meta's research blog said Muse Wonder and Muse Wonder Full deliver frontier performance on personal and professional agentic tasks, including work in law, finance, health and coding. The company built the models around two named capabilities: Active Memory, for carrying important context about the user across conversations, and Active Focus, for sustained effort on complex goals, including recovering from errors and asking for clarification rather than guessing. Meta's own announcement names navigating insurance claims, understanding leases, making sense of finances, and building apps or games as example tasks the models are meant to handle.
Muse Wonder is available immediately through the Meta Model API and Muse Code, while the more capable Muse Wonder Full is reserved for Muse Code subscribers rather than being served through the general API. Both are coming to the Muse platform for subscribers. TechCrunch reported that Meta is pricing Muse Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens, and that the models were trained using Meta's Prometheus data centre in central Ohio, which came online earlier in 2026.
Key details
| Detail | What Meta and TechCrunch reported |
|---|---|
| Release date | 12 October 2026 |
| Models | Muse Wonder and Muse Wonder Full |
| Headline capability | Active Memory (context persists across conversations) and Active Focus (sustained effort on complex goals) |
| Access | Muse Wonder via Meta Model API and Muse Code; Muse Wonder Full for Muse Code subscribers only |
| Pricing (Muse Wonder, via API) | $4 per million input tokens, $20 per million output tokens (TechCrunch) |
| Training infrastructure | Meta's Prometheus data centre, central Ohio (TechCrunch) |
Why it matters
Most business use of AI chat tools today resets each session: you re-explain your business, your preferences and the task every time you open a new conversation. Meta's pitch for Active Memory is that the model carries that context forward itself. Meta also says the Muse Wonder models were developed under its Superintelligence Scaling Framework with its strongest safeguards to date against CBRN and cyber misuse, and calls them its most aligned models to date, with a full safety evaluation report to follow. TechCrunch reported that Meta says the two models achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra, and score higher than Fable, Opus or Astra on what Meta calls MemoryBench, which is Meta's own comparison, not an independent one.
What this means for businesses
A model that remembers your business's context sounds convenient, but it also means more of your business's information is being stored and carried forward somewhere outside your own systems. Before adopting any tool built on Muse Wonder or a similar memory-carrying model, a small or medium business should check:
- What exactly gets stored as part of its memory feature, for how long, and whether you can review or delete it.
- Whether the pricing model (per-token, for both input and output) fits how heavily your team would actually use it, since output tokens are charged separately and can add up quickly in longer conversations.
- Whether the more capable version your team actually needs (Muse Wonder Full) is the one locked behind a separate subscription, not the one available through the general API.
- Whether Meta's own safety evaluation report, once published, addresses the specific use case you have in mind, rather than relying on launch-day claims alone.
If your business is weighing up how an AI agent with persistent memory would fit into client work in professional services, it is worth treating the decision the same way as any other new business system: understand what data goes in, who can see it, and what it costs at the volume you would actually use. Our IT consulting service can help you work through that before you commit a team to a new tool, and our earlier look at AI agents, permissions and cost for a small business covers the same questions for agent tools generally. If you want a second opinion on whether a memory-carrying AI agent fits your business before you sign up, get in touch for a quote.
Frequently asked questions
What did Meta actually announce on 12 October 2026?
Meta announced Muse Wonder and Muse Wonder Full, which it describes as its most capable models to date, with Active Memory for carrying context across conversations and Active Focus for sustained effort on complex tasks.
How much does Muse Wonder cost to use?
TechCrunch reported that Meta is pricing Muse Wonder through the Meta Model API at $4 per million input tokens and $20 per million output tokens; Muse Wonder Full is only available to Muse Code subscribers, not priced through the API.
What can these models actually do?
Meta's announcement lists navigating insurance claims, understanding leases, making sense of finances, and building apps or games as example tasks, alongside stronger performance on coding and professional-domain work such as law, finance and health.
Is Muse Wonder Full available to everyone?
No. Meta says Muse Wonder is available immediately through the Meta Model API and Muse Code, while Muse Wonder Full is reserved for Muse Code subscribers rather than being served through the general API.
Has Meta published independent safety results for these models?
Meta says the models were developed under its Superintelligence Scaling Framework with its strongest safeguards to date and calls them its most aligned models yet, but says a full safety evaluation report will follow, so the detailed results are not yet public.
Where were the Muse Wonder models trained?
TechCrunch reported that Meta trained the models using its Prometheus data centre in central Ohio, which came online earlier in 2026.
Key takeaways
- Meta announced Muse Wonder and Muse Wonder Full on 12 October 2026, built around Active Memory and Active Focus for agentic work.
- Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens through the Meta Model API, per TechCrunch; Muse Wonder Full is Muse Code subscriber-only.
- Meta names insurance, leases, personal finance and app or game building as example use cases, plus professional-domain and coding work.
- Meta's safety claims are launch-day statements; a full safety evaluation report has not yet been published.
- Before adopting a memory-carrying AI tool, check what it stores, for how long, and what it actually costs at your team's real usage volume.