Meta unveils Muse Wonder models with persistent memory and split API access
The AMW Read
The release meaningfully updates Meta's foundation-model offering through persistent memory and differentiated access, but self-reported benchmarks do not resolve competitive performance debates.
Meta unveils Muse Wonder models with persistent memory and split API access
Meta Superintelligence Labs has introduced Muse Wonder and Muse Wonder Full, larger models than its lightweight Muse Spark, according to ZDNet Korea. Meta says they improve coding and complex task handling, with both powering Muse Agent. Muse Wonder is available through Meta's model API at $4 per million input tokens and $20 per million output tokens. Wonder Full is restricted to Muse Code subscribers and has no API offering. Meta's own model-card tests report scores comparable to Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra; these are company-reported results, not independent confirmation.
The foundation-model competitive signal is Meta's emphasis on persistent personal context alongside task performance. Its active-memory capability is intended to retain and combine user-provided information across work lasting months. Meta also reports that Wonder outscored competing models on MemoryBench in its own testing. If sustained in actual use, reliable memory could make assistants more useful for continuing projects and make retained context a source of product differentiation. The reported scores alone do not establish that advantage.
Builders should evaluate the API-accessible Wonder separately from subscription-only Wonder Full: the access split means the two models cannot be treated as interchangeable integration options. A concrete pilot should test recall accuracy over extended tasks, handling of changed user information, and total token costs at the published rates. For investors, repeat use driven by dependable memory would be stronger evidence of differentiation than launch benchmarks alone.
