Meta says Muse Wonder can match the best models. The number to check is what you'll pay.
Meta's new agent models rank near the top of one popular index, by its own count. The paid tiers and the missing API price are where the launch gets real.
Meta released two new models on Monday, and it is calling them “our most capable models to date.” They are Muse Wonder and Muse Wonder Full. The benchmark charts are the part everyone will quote. I think the more useful question is who gets to use them and what it costs them.
What Meta is claiming
Meta says the larger model, Muse Wonder Full, outperforms Anthropic’s second-best model, Claude Fable 5.1, and sits just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. That index is a composite score many people use to compare models. Meta also says both models beat OpenAI’s GPT-6 Astra on some specialised tests, including legal document work and building software.
These are Meta’s own numbers from its own testing. The Verge says the results come from internal testing and that it is unclear how the models compare with Google’s Gemini 4 Argon. WIRED adds the caution that benchmarks don’t always predict real-world performance. I’d treat the ranking as a claim until an outside group reruns it.
Two features aimed at the “gives up halfway” problem
Meta highlights two new features, Active Memory and Active Focus. Active Memory is meant to keep track of what matters to you. Active Focus is meant to keep the agent working toward a goal.
The clearest statement of the goal came from Alexandr Wang, Meta’s AI chief, in a CNBC interview. “We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time,” he said. That is a sensible target. Many agents fail by asking permission every few steps. But Meta has not published how these models were trained, so the claim about behaviour is one I can’t check from the launch material.
The part to check: who pays
The consumer Muse app stays free to start. Its paid plans are $20 and $100 a month, CNBC reports. Muse Wonder Full is available to Muse Code subscribers. Muse Wonder is available through the Meta Model API for developers.
Wang told CNBC the API price will be higher than the standard subscriptions because the model is larger, and that it will still be priced “meaningfully below” the leading models from OpenAI and Anthropic. He did not give a number. So right now the developer price is unknown, and I would not repeat any comparison until Meta puts a figure on its pricing page.
Meta also says the model “offers our strongest safeguards to date” against cyber misuse. CNBC notes those protections only come with the paid model for now.
Privacy complaints, and what Meta says about them
CNBC reports that some users posted on social media that the Muse agent read personal messages without permission. Meta disputes that. Wang’s explanation was that people sometimes forget they authorised a conversation earlier in a chat. Meta says it is educating users on this. I can’t verify either account from here, so I’d call it an open question. If you give an agent access to your messages, read what it’s allowed to see before you rely on it.
What I’d watch
- The full safety report. Meta’s post says it “will be available soon.”
- When Muse Wonder reaches app subscribers. Meta says they get it first, with no date.
- The developer price. That number will tell you more about Meta’s plans than any benchmark.
- An outside ranking. If Artificial Analysis or another lab confirms the placement, the claim gets much stronger.
My own read: the benchmark chart is the least interesting part of this launch. Meta is betting that people will pay $20 or $100 a month for an agent that finishes tasks. Whether that bet pays off depends on how the agent behaves on a Tuesday afternoon with your email open, and that’s not something a launch post can show. I haven’t used the models, so this is a read of the announcement, not a test.
For context on what happens when agents touch shopping and payments, see our earlier post on Amazon locking AI shopping agents out of its checkout. For the pricing side of the wider model race, see OpenAI’s price cuts.
Sources
- Meta, “Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents,” October 12, 2026: about.fb.com/news/2026/10/introducing-muse-wonder-models
- Meta AI Research, “Muse Wonder,” October 12, 2026: research.meta.ai/blog/muse-wonder. The linked model card PDF could not be read for this post.
- Jonathan Vanian, “Meta debuts Muse Wonder AI model in test of users’ willingness to pay,” CNBC, October 12, 2026: cnbc.com/2026/10/12/meta-muse-wonder-ai-model-agent.html
- Will Knight, “Muse Wonder Puts Meta Back at the AI Frontier,” WIRED, October 12, 2026: wired.com/story/muse-wonder-puts-meta-back-at-the-ai-frontier
- Robert Hart, “Meta has some Wonder Full new AI models,” The Verge, October 12, 2026: theverge.com/ai-artificial-intelligence/1009479/meta-muse-wonder-full-ai-model