Meta launches Muse Wonder and Wonder Full with a million tokens

What's new

On October 12, 2026, Meta unveiled two new AI models: Muse Wonder and Muse Wonder Full. They are far larger and more advanced than Meta's lightweight model Muse Spark. Both are proprietary, so they cannot be downloaded and run freely like open models.

Muse Wonder is available through the Meta Model API (access for developers and businesses) and in the Muse Code tool. Muse Wonder Full is not available through the API and is only for Muse Code subscribers. Wonder Full is also expected to come to Meta's Muse agent for subscribers later, but no date has been announced. Both models have a context window of one million tokens, meaning room to hold very large amounts of text or code in mind at once.

What's smart

Meta highlights two features. Active Memory remembers important information across conversations and brings it up when it is relevant. Active Focus is meant to make the models stick with difficult tasks, try other approaches, fix their own mistakes, ask for clarification and check actions before carrying them out.

Muse Code can be used on Mac as a terminal-based agent, meaning a program that works independently through text commands, and it is also accessible via iPad.

On data: Data sent to the API is not used for training by default. Muse Code is set to Contributor mode by default, which allows use for improving the model. Zero data retention, where data is not stored, is available after approval.

Cheaper or better?

Muse Code costs $50 a month (Power Usage) or $200 a month (Maximum Usage). Both tiers include access to Wonder Full. For Muse Wonder in the API, the price is $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Web search costs $2.50 per 1,000 requests. That matches the price of Claude Opus 5.5 and is below GPT-6 Astra. By comparison, Claude Sonnet 5.5 costs half as much: $2 for input and $10 for output per million tokens. Fello AI points out that Meta's pricing is not noticeably below the competition.

In Meta's own tests, the results look good. On MemoryBench, Wonder scored 85.6 percent and Wonder Full 85.5 percent, versus 81.5 for Claude Opus 5.5 and 81.4 for GPT-6 Astra. On long coding tasks (SWE-Marathon), Wonder scored 53.1 percent and Wonder Full 57.5 percent, versus 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. These are the manufacturer's own figures. In Meta's own overview, Wonder Full wins 7 of 14 rows, including on legal tasks (13.1 versus 4.2 for Opus 5.5), while Opus 5.5 leads in 6.

Independent measurements are more subdued. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scored 58, Wonder Full 56 and Wonder 52. Wonder Full was on par with Claude Sonnet 5.5 and above GPT-6 Astra and Gemini 4 Argon, while Wonder was one point below GPT-6 Astra. Fello AI assesses that Wonder's advantage is narrow and concentrated in long coding tasks and legal work, not broad superiority.

What it's good at

The models are aimed first and foremost at software development. Wonder Full scored 79.0 on DeepSWE 1.1, a software development test, and performs strongly on long coding tasks where the AI has to work through many steps.

That could matter to developers and companies that want an agent for larger programming tasks. The same goes for legal work, where Meta itself shows a large lead, and for tasks where it helps that the model remembers earlier conversations.

For general users, it is important to know that Wonder Full requires a Muse Code subscription starting at $50 a month. The sources do not say whether the models are available in Denmark or the EU.