Muse Wonder is Meta’s first frontier model, meaning a large model built to compete with the most capable systems on the market. Meta’s Superintelligence Lab released it on October 12, 2026, alongside Muse Wonder Full, in what TechCrunch calls the company’s first attempt at a frontier model. It follows Meta’s lightweight Muse Spark model and the Muse agent.
How does Muse Wonder benchmark against GPT-6 and other top models?
According to Meta’s own testing, presented in the model card, Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra. The standout result is on MemoryBench (a benchmark that tests AI systems’ ability to learn from accumulated user feedback across sessions, not just read long documents in a single conversation). Muse Wonder now scores higher on MemoryBench than Fable, Opus, or Astra, per Meta’s testing.
Meta’s blog post describes the two models as “the next step on our scaling ladder and a major leap toward personal superintelligence.” Research scientist Kavya Kopparapu said in a video accompanying the launch that the team put “a tremendous amount of effort into making sure your model can remember what’s important to you over days, weeks, months, and beyond.”
For context on what MemoryBench measures: the benchmark was proposed in a paper on arxiv.org specifically because existing memory benchmarks mostly tested reading comprehension on long inputs, rather than the ability to build memory from ongoing user interactions. In the paper’s own experiments, the effectiveness and efficiency of state-of-the-art baselines were “far from satisfying.”
What is “active memory” and how does it differ from standard AI responses?
Meta calls this retention capability “active memory,” focusing on synthesizing and recalling information a user provides across multiple sessions over time. Where a context window covers only the text loaded into a single conversation, active memory aims to keep what a user has shared over days, weeks, and longer.
The MemoryBench benchmark tests this distinction directly. According to the arxiv.org paper, it evaluates how well AI systems learn from accumulated user feedback in service time, covering multiple domains, languages, and task types. The paper’s authors argue that scaling up data, parameters, and test-time computation, the mainstream ways to improve these systems, is nearing its limits as high-quality data runs short and extra compute brings smaller gains, which is why memory and continual learning have become a popular research direction.
How is Muse Wonder priced, and who can access Wonder Full?
Accessed through the Meta Model API, Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens. Tokens are the chunks of text the model reads and generates. TechCrunch reports this is roughly equivalent in cost to GPT-6 Sol or Opus 5.5.
Muse Wonder Full, the larger variant, will not be available through the API. It will only be accessible to Muse Code subscribers. Both models will also be used to power the Muse agent.
Update, October 13, 2026: The Verge reports that Muse Wonder and Muse Wonder Full are available to Muse Code subscribers starting October 12, and that Wonder can also be accessed through the Meta Model API. Meta says it eventually plans to power the Muse agent with Wonder, starting with subscribers, but has not shared specifics or a timeline. The Verge also notes it is unclear how the models compare with Google’s Gemini 4 Argon.
The models were trained at Meta’s Prometheus data center in central Ohio, which came online earlier this year, according to TechCrunch. This release is Meta’s first frontier model since Scale AI founder Alexandr Wang restructured Meta’s AI efforts in August of last year. Since that restructuring, Meta had released its lightweight Muse Spark model and the Muse agent, but had avoided the kind of state-of-the-art model releases OpenAI and Anthropic are known for, TechCrunch notes. Where Muse Spark aimed for low latency at smaller scale, the Wonder models are larger and described as more capable, with higher benchmark scores and better results on complex tasks such as coding, per TechCrunch.
Before building on Muse Wonder’s active memory
- Confirm whether your use case needs multi-session persistent memory or standard single-session long context. They are different capabilities with different benchmark criteria.
- Check access requirements: Wonder Full is restricted to Muse Code subscribers and unavailable via the standard API.
- Compare per-token pricing ($4 input / $20 output per million tokens) against your current model costs at expected usage volume.
- Review the MemoryBench paper (arxiv.org/abs/2510.17281) to understand what the benchmark actually measures and where current systems fall short.
- Run your own multi-session tests before assuming Meta’s benchmark scores reflect production behavior in your specific domain or language.
- Track independent evaluations of active memory as they appear, since the published scores come from Meta’s own testing.
What this means for AI-search visibility
In Hingewise’s view, an assistant that remembers user preferences works differently as a discovery channel from one that starts each conversation from scratch, and that is the main implication of active memory for brands that want to appear in AI-generated answers.
Suppose a user tells the assistant they prefer software tools with a free tier, and that preference is kept for weeks. Later recommendations could then be filtered through it, favoring brands that already fit. Meta has not said how stored memories shape recommendations, so this is a scenario, not a documented behavior.
If memory works that way, visibility in AI answers becomes more cumulative: what an assistant learns about a user’s needs early could carry into later sessions. Brands that pick up unfavorable associations early may be more exposed than brands that simply have not been mentioned yet.
Two important caveats. First, Meta’s MemoryBench scores come from Meta’s own testing, not independent third-party evaluation. How well Wonder’s active memory holds up in everyday use is unknown until outside evaluations appear. Second, Wonder Full is limited to Muse Code subscribers and is not served through the API, although Meta says it plans to use Wonder to power the Muse agent, starting with subscribers.
Hingewise’s assessment: persistent memory in frontier models could raise the stakes for early brand associations in AI-assisted conversations. The result worth watching is less Meta’s own MemoryBench score than how independent evaluators rate multi-session memory over the coming months.
Sources: TechCrunch AI, October 12, 2026; MemoryBench paper, arxiv.org; The Verge, October 12, 2026
Lam Nguyen · Hingewise
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.