Meta launches Muse Wonder and Muse Wonder Full with memory

What's new

Meta has launched two new AI models: Muse Wonder and Muse Wonder Full. They were unveiled on October 12, 2026. Meta calls them its most capable models to date and says they are built on top of Muse Spark. They are aimed at agents, meaning AI that can handle tasks for you on its own instead of just answering questions. That includes law, finance, health and coding.

Muse Wonder is available through Meta's API (the access developers use to build AI into their own products) and in the coding tool Muse Code. According to Fello AI, it is also on OpenRouter and is coming to the Muse app's paid tiers at $20, $100 and $200 a month. Meta itself says the model will eventually power more of the Muse app, first for subscribers, but without a timeline. Muse Wonder Full is only for paying Muse Code subscribers, at $50 to $200 a month, and cannot be bought through the API. Both have room for 1 million tokens (chunks of text) at a time, and according to one review they can accept text, images, video and audio. The models are closed, so no open weights have been released.

What's smart

Meta highlights two features. Active Memory lets the model carry relevant information across conversations, so you don't have to explain everything from scratch. Active Focus is meant to keep the model working toward the goal when things get hard: it should try other approaches and correct its own mistakes. It should also ask questions when necessary and check with you before doing anything important.

Meta says the models are trained not to share private information without permission and to refuse harmful requests. They should still be usable for finding bugs in your own code. A full safety report is promised "soon."

Cheaper or better?

In the API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens. That is the same price as Anthropic's Claude Opus 5.5 and cheaper than OpenAI's GPT-6 Astra, which costs $10 and $50. There is no API price for Muse Wonder Full, since it isn't sold there.

On performance, the picture is mixed. In Meta's own tests, Muse Wonder Full scored 59.1 on the Personal Superintelligence Index versus 55.2 for Opus 5.5 and 54.5 for Astra. On the coding test SWE-Marathon, it scored 57.5 versus 46.3 and 42.5. On Harvey's legal test, it scored 13.1 versus 8.6 for Astra and 4.2 for Opus 5.5. On the memory test MemoryBench, both Wonder models scored around 85.5 versus 81.5 for Opus 5.5.

But the numbers are Meta's own. Meta has itself said that the coding comparison was not controlled and that the model ran at maximum effort. According to Fello AI, Muse Wonder Full won 7 of 14 rows in Meta's table, while Opus 5.5 won 6. The independent evaluator Artificial Analysis has Opus 5.5 on top with 58 points, Muse Wonder Full at 56 and Muse Wonder at 52. On Threads, Mark Zuckerberg calls the launch a big leap toward personal superintelligence for everyone.

What it's good for

According to Meta, the models are built for personal and professional agent tasks: coding, legal work, finance and health. The memory can be especially useful if you work on the same task over several days. Muse Wonder suits companies and developers who want a powerful model at the same price as Opus 5.5 but cheaper than Astra. Muse Wonder Full is for programmers willing to pay for a Muse Code subscription.

For ordinary Danes, it is unclear when or whether the Muse app will become available. The sources do not say whether the models can be used in Denmark or the EU.