TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on Oct. 12, describing them as its first frontier models. Meta says their benchmark performance is comparable to Fable 5.1 and GPT-6 Astra, and highlights improved performance on a memory benchmark; independent results and broader product details were not included in the report.
Meta’s Superintelligence Lab released two AI models, Muse Wonder and Muse Wonder Full, on Oct. 12, calling them the company’s first frontier models. Meta says they are larger and more capable than its earlier Muse Spark model, and that their benchmark scores compare with Fable 5.1 and GPT-6 Astra; the release marks a move into the high-capability model tier associated with competitors such as OpenAI and Anthropic.
The models are positioned above Muse Spark, which Meta introduced as a lightweight system focused on low latency. In its model card, Meta reports that Wonder and Wonder Full achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra, with improved results on complex work such as coding. Those comparisons come from Meta’s own testing; the report does not provide independent evaluations or enough detail to judge how the systems perform across all uses.
Muse Wonder will be offered through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Meta says Muse Wonder Full will not be served through the API and will instead be available only to Muse Code subscribers. Both models are also set to power Meta’s Muse agent. The announcement does not specify when access will open to all intended users or whether the subscription model will change.
Meta also emphasized “active memory”: its term for retaining and synthesizing information supplied by a user over time. The company says Muse Wonder scores higher on MemoryBench than Fable, Opus and Astra. The benchmark is intended to test continual learning and memory construction, according to the report. Meta says the models were trained at its Prometheus data center in central Ohio, which came online earlier this year.
Meta Enters the Frontier Model Race
The release is a meaningful change in Meta’s AI product strategy. After launching the lighter Muse Spark and its Muse agent, the company had not released a model positioned as a direct state-of-the-art competitor. Wonder brings Meta into that market with an API offering, while Wonder Full is tied to a coding subscription rather than broad API access.
Its emphasis on memory also points to how Meta hopes its models will be used: as assistants that can act over longer periods and make use of information a person has shared previously. If that capability works well in practice, it could affect the usefulness of personal assistants and coding tools. But Meta’s benchmark claims alone do not establish how reliably the models remember information in ordinary use, or how users’ stored information will be managed.
Pricing may matter to developers choosing among models. Meta’s reported API rates place Wonder at $4 per million input tokens and $20 per million output tokens, which TechCrunch described as roughly equivalent to GPT-6 Sol or Opus 5.5. The price comparison does not by itself show which model offers better value; that depends on performance, usage and the tasks being run.
As an affiliate, we earn on qualifying purchases.
From Muse Spark to Wonder
The announcement follows a reorganization of Meta’s AI work by Alexandr Wang, founder of Scale AI, in August 2025, according to the source report. Since then, Meta had released Muse Spark and the Muse agent but had not made a frontier-model launch of the kind associated with leading AI labs. Wonder is the first such release in that period, based on the report’s account.
Meta’s blog post describes the pair as the next step on a “scaling ladder.” The company says the larger models are intended to handle more complex tasks than Spark, while the planned use in Muse’s agent links model capability to assistant features. The report identifies Prometheus in Ohio as the training site but does not give details about the models’ training data, compute scale or development process.
“Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence.”
— Meta blog post
As an affiliate, we earn on qualifying purchases.
Independent Testing Remains Pending
The reported comparisons and MemoryBench results are Meta’s own test findings. The source does not cite independent benchmark testing, give the full scores, or describe the testing conditions, so the scale and practical meaning of the reported advantages remain uncertain. Comparable benchmark scores also do not establish that models perform equally well across different tasks or real-world settings.
Further details about availability, safety evaluations and memory controls are not provided in the report. It is not clear how users can inspect, correct or delete information retained through active memory, or how consistently the models will recall relevant details across long periods. The scope of Muse Wonder Full’s subscriber access and the timing of wider rollout also remain unspecified.
As an affiliate, we earn on qualifying purchases.
API and Muse Rollout Details
The next developments to watch are access to Muse Wonder through Meta’s API and the deployment of both models in the Muse agent. Meta has not given a rollout schedule in the source report, nor has it announced when further technical or safety documentation will be available. More information about Wonder Full access for Muse Code subscribers is also pending.
Independent evaluations could help clarify how the models compare with rival systems on coding, memory and other tasks. Until such results and additional product details are available, the strongest confirmed points are the models’ release, their stated access routes and prices, and Meta’s own claims about benchmark performance.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta announce?
Meta’s Superintelligence Lab released Muse Wonder and Muse Wonder Full on Oct. 12, 2026. The company describes them as its first frontier models.
How can users access the new models?
Muse Wonder is planned for the Meta Model API. Meta says Muse Wonder Full will be available only to Muse Code subscribers and will not be served through the API. Both models are also intended to power the Muse agent.
What does Muse Wonder cost?
Meta lists Wonder at $4 per million input tokens and $20 per million output tokens through the Meta Model API. The report gives no API price for Wonder Full because it will not be offered there.
Are the benchmark claims independently verified?
Not in the source report. The comparisons with Fable 5.1 and GPT-6 Astra, along with the MemoryBench result, are attributed to Meta’s testing; independent evaluations were not cited.
What does Meta mean by active memory?
Meta uses “active memory” to describe a model’s ability to retain and synthesize information supplied by a user over time. The report does not detail how users can manage or delete that retained information.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
