Meta Ships Its First Frontier Model. The Best Version Is Not on the API.
Fourteen months after Zuckerberg blew up Meta's AI org and handed it to Alexandr Wang, the lab shipped Muse Wonder and Muse Wonder Full on Monday and called them frontier. The number everyone passed around within the hour: Wonder Full lands at 56 on the Artificial Analysis index, second only to Opus 5.5 at 58, up from 48 for Muse Spark 1.3 six weeks ago. Meta stock closed up more than three percent, Hacker News put it at 450 points, and the pub interview with Zuckerberg and Wang had 29,000 views by evening. Wonder costs 4 dollars per million input tokens and 20 per million output on the Meta Model API, the same bracket as Opus 5.5 and GPT-6 Sol. Context is one million tokens.
The pitch is agents, not chat. Meta's two words are Active Memory, carrying what matters about you across sessions and pulling it back when relevant, and Active Focus, which is the model trying alternative approaches, working through blockers, and checking in before an important action. The blog leans on an internal MemoryBench where Wonder beats Fable, Opus and Astra, on FocusBench for persistence across housing, travel, IT and finance tasks, and on SWE-Marathon for multi-day software projects, where Wonder Full posts 57.5. Meta also published a Personal Superintelligence Index, a composite of seven evals, with Wonder Full on top at 59.1. Read the model card before quoting that: two of the seven evals are Meta-internal, the weights are not published, the SWE-Marathon row is explicitly not a controlled common-harness comparison, and Meta says its runs of competitor models are best-effort.
The catch for builders is the split. Wonder Full, the one that posts the headline numbers, is only available to Muse Code subscribers and is not served through the API at all. The API model is plain Wonder, which on Meta's own index sits fourth of five, basically tied with Astra and a hair behind Opus. Zuckerberg told the FT the strongest version will be sold directly to enterprises and will not be open-sourced. So the company that built its AI brand on Llama now runs a closed frontier model, gates the best tier behind its own coding harness, and prices the open tier like Anthropic. That is the Anthropic playbook executed by a company with three billion users and a gigawatt data center in Ohio.
Two things to watch. First, whether the memory claim survives contact: the model card recommends human-in-the-loop confirmation for irreversible actions, which is the same containment language every lab adopted this weekend. Second, whether Wonder Full ever reaches the API. Until it does, every third-party comparison is of the second-best model, and Meta gets to keep the real one for its own agent.
Blog: https://research.meta.ai/blog/muse-wonder
Model card: https://research.meta.ai/static/muse-wonder-model-card
← Back to all articles
The pitch is agents, not chat. Meta's two words are Active Memory, carrying what matters about you across sessions and pulling it back when relevant, and Active Focus, which is the model trying alternative approaches, working through blockers, and checking in before an important action. The blog leans on an internal MemoryBench where Wonder beats Fable, Opus and Astra, on FocusBench for persistence across housing, travel, IT and finance tasks, and on SWE-Marathon for multi-day software projects, where Wonder Full posts 57.5. Meta also published a Personal Superintelligence Index, a composite of seven evals, with Wonder Full on top at 59.1. Read the model card before quoting that: two of the seven evals are Meta-internal, the weights are not published, the SWE-Marathon row is explicitly not a controlled common-harness comparison, and Meta says its runs of competitor models are best-effort.
The catch for builders is the split. Wonder Full, the one that posts the headline numbers, is only available to Muse Code subscribers and is not served through the API at all. The API model is plain Wonder, which on Meta's own index sits fourth of five, basically tied with Astra and a hair behind Opus. Zuckerberg told the FT the strongest version will be sold directly to enterprises and will not be open-sourced. So the company that built its AI brand on Llama now runs a closed frontier model, gates the best tier behind its own coding harness, and prices the open tier like Anthropic. That is the Anthropic playbook executed by a company with three billion users and a gigawatt data center in Ohio.
Two things to watch. First, whether the memory claim survives contact: the model card recommends human-in-the-loop confirmation for irreversible actions, which is the same containment language every lab adopted this weekend. Second, whether Wonder Full ever reaches the API. Until it does, every third-party comparison is of the second-best model, and Meta gets to keep the real one for its own agent.
Blog: https://research.meta.ai/blog/muse-wonder
Model card: https://research.meta.ai/static/muse-wonder-model-card
Comments