New to this? Read it in simple words
- Meta released two new AI models, Muse Wonder and Muse Wonder Full, on October 12.
- Developers can use Muse Wonder for $4 per million input tokens and $20 per million output tokens.
- An outside test scores Muse Wonder Full 56, two points behind Claude Opus 5.5.
- It is weaker on facts, and costs about three times as much per task as Meta’s last model.
- Frontier model
- One of the most capable AI models available at a given time.
- Token
- A small piece of text, about part of a word, that AI models read and write.
- API
- A way for other programs to use a service, here to send work to Meta’s model.
Meta Superintelligence Labs began rebuilding Meta’s AI work fourteen months ago. Its first model, Muse Spark, was small and fast. The new Muse Wonder series is Meta’s first attempt at a model that competes with the best.
Two models, two ways in
Meta released Muse Wonder and Muse Wonder Full on Monday, October 12. It calls them its “most capable models to date”, trained at its Prometheus data centre.
Muse Wonder is available through the Meta Model API and in Muse Code, Meta’s coding agent. Through the API, it costs $4 per million input tokens and $20 per million output tokens.
Muse Wonder Full is not offered through the API. It is only for Muse Code subscribers on the $50 or $200 monthly plans. Meta says Muse Wonder will soon power its Muse agent app as well, first for paying subscribers.
Artificial Analysis Intelligence Index scores at each model’s highest setting, as of October 12, 2026.
What Meta says the models can do
Meta says both models reach “state-of-the-art performance on personal and professional agentic tasks”, such as an insurance claim or a lease. These results are Meta’s own tests.
Meta names two new abilities. “Active Memory” carries what matters to a user across conversations. “Active Focus” keeps working on a hard task, tries other ways and asks before important steps.
Meta says the models have its strongest protections yet against chemical, biological, nuclear and cyber misuse. A full safety report will come later, the company says.
What an outside test found
Artificial Analysis, an independent group that tests AI models, scores Muse Wonder Full at 56 on its Intelligence Index. That ties it with Claude Sonnet 5.5, puts it ahead of Claude Fable 5.1 and GPT-6 Astra at 53, and two points behind Claude Opus 5.5 at 58.
Muse Wonder scores 52, and Meta’s earlier Muse Spark 1.3 scored 48. The biggest gains are in coding and agent tasks, The Decoder reports from the test data.
Factual knowledge is weaker. On a knowledge test, Muse Wonder Full scores 30, against 46 for Claude Opus 5.5. Running the whole index costs $4.52 per task with Muse Wonder, almost three times as much as Muse Spark 1.3.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 12, 2026Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents Meta
- 2
- 3
- 4
- 5Research · October 12, 2026Meta debuts Muse Wonder AI model in latest test of users’ willingness to pay CNBC
- 6Research · October 12, 2026Meta says it’s caught up to the frontier with its latest Wonder AI models Engadget
- 7Research · October 12, 2026Meta’s new frontier model Muse Wonder Full breaks into the AI top tier The Decoder
The “state-of-the-art” results and the two new abilities are Meta’s own claims. The scores are Artificial Analysis’s first results of October 12, and they can change as it tests more.
Meta’s new models launch close to the AI frontier in an outside test, though they are weaker on facts and cost more per task. How we rate the mood