Meta's Superintelligence Lab released two new state-of-the-art models on Monday, Muse Wonder and Muse Wonder Full. This marks the company's first attempt at a frontier model, signaling a major push into advanced AI development.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Meta's Superintelligence Lab has released Muse Wonder and Muse Wonder Full, Meta's first frontier models since Alexandr Wang restructured its AI work. The core selling point is "active memory," but all benchmark comparisons and pricing benchmarks come from Meta's own testing and statements, with no third-party independent verification yet.
Meta's Superintelligence Lab released two models, Muse Wonder and Muse Wonder Full, on Monday, the company's first attempt at frontier models. The release time was 10:56 a.m. Pacific Time on October 12, 2026.
The timing of this release is notable: it is Meta's first frontier model release since Scale AI founder Alexandr Wang restructured Meta's AI work last August. Since then Meta had launched the lightweight Muse Spark model and the Muse agent, but had avoided the kind of industry-leading model releases from OpenAI and Anthropic.
Unlike Muse Spark, which is positioned as a low-latency lightweight model, the Wonder series models are larger and said to be more capable, with higher benchmark scores and better performance on complex tasks such as coding. In a blog post, Meta called the two models "the next step on our scaling ladder, a major leap toward personal superintelligence."
Meta particularly emphasizes the new models' ability to retain and synthesize information given by users, calling this feature "active memory." This is consistent with its interest in developing personal assistants that can carry out tasks over the long term.
Research scientist Kavya Kopparapu said in an accompanying launch video that the team put a great deal of effort into ensuring the models can remember what matters to users over days, weeks, months and even longer.
Meta highlights the MemoryBench benchmark, used to test continuous learning and memory-building frameworks, and says that in its testing Muse Wonder scores higher on that benchmark than Fable, Opus or Astra. However, this claim is only Meta's own statement; the report does not give specific benchmark score figures, only relative comparisons.
When accessed through the Meta Model API, Wonder is priced at $4.00 per million input tokens and $20.00 per million output tokens, roughly comparable in cost to GPT-6 Sol or Opus 5.5.
Notably, Wonder Full is available only to Muse Code subscribers and will not be served through the API. This means external developers cannot test Wonder Full's true capabilities through a public API.
Both models will be used to power the Muse agent.
According to tests Meta published in the model card, the benchmark scores of Wonder and Wonder Full are comparable to Fable 5.1 and GPT-6 Astra. All benchmark comparisons in the report, including those with Fable 5.1, GPT-6 Astra and MemoryBench, come from Meta's own testing, with no third-party independent verification seen.
The report does not give specific benchmark score figures, only relative comparisons. Therefore, whether the models truly reach a "frontier" level can currently only be conveyed through Meta's positioning, lacking external evaluation support.
Likewise, the claim that Muse Wonder leads Fable, Opus or Astra on MemoryBench is only Meta's own statement. The actual effect of "active memory" over days, weeks and months also lacks external evaluation.
According to the blog post, these models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year.
This training detail comes only from Meta's blog, with no independent source to corroborate it.
From the verifiable information, Meta did release Muse Wonder and Muse Wonder Full, the pricing and distribution methods are clear, and Wonder Full is limited to subscribers. But all capability benchmarks—comparable to Fable 5.1 and GPT-6 Astra, leading on MemoryBench—come from Meta's own testing and statements, with no third-party independent verification.
This means readers can treat "Meta released its first frontier model" as a confirmed fact, but treat "the model reaches frontier level" as Meta's self-positioning. Wonder Full is aimed only at subscribers, and its true capabilities cannot be externally tested through a public API, further increasing the difficulty of verification.
All facts in this article come from TechCrunch's report on Meta's release. Among them, the release event, pricing, distribution methods and training location are facts confirmed by the report; the benchmark comparisons and the effects of "active memory" are Meta's own statements, without third-party independent verification. The report does not give specific benchmark score figures, only relative comparisons.
Meta has finally delivered its first frontier model, but for now this report card is signed only by Meta itself. Active memory is a differentiation direction worth watching, but until third-party evaluations appear, its true level remains to be tested.
Primary report
Primary source