Meta Debuts Muse Wonder, Its First Frontier Model
Meta Debuts Muse Wonder, Its First Frontier Model
摘要: Meta Launches Muse Wonder, Its First Frontier AI Model, at $4 per Million Tokens
Meta’s Muse Wonder Models
Meta now prices its AI lik
正文
Meta Launches Muse Wonder, Its First Frontier AI Model, at $4 per Million Tokens
Meta’s Muse Wonder Models
Meta now prices its AI like a frontier lab. Meta Superintelligence Labs
launched Muse Wonder on Monday
, the company’s first frontier-class model, at $4 per million input tokens and $20 per million output tokens through the Meta Model API, while keeping a stronger sibling, Muse Wonder Full, inside its paid Muse Code coding tool.
Muse Wonder art. Image credit: Meta
Key Takeaways
Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 and a one-million-token context window.
Muse Wonder Full is not sold through the API; developers can use it only on Muse Code’s Power Usage and Maximum Usage plans.
Meta’s own tables put both models ahead on a memory benchmark, while the third-party Artificial Analysis Index ranks Wonder Full at 56, behind Claude Opus 5.5 at 58.
Two Models, Two Doors
Muse Wonder runs under the model ID muse-wonder-1.0. Developers who already build on the OpenAI SDK can switch by changing one base URL. Cached input costs $0.20 per million tokens, and the context window holds one million tokens.
Muse Wonder Full takes a different route. Meta offers it only on two Muse Code plans: Power Usage and a new Maximum Usage tier, which carries 80 times the limit of the entry-level Everyday Usage plan. Those plans cost $50 and $200 a month. Meta has not published a per-token price for Wonder Full.
Built to Remember
Meta aims the series at long-running agents, coding and computer use. A feature it calls active memory carries context across conversations and keeps goals, constraints and earlier decisions in view as work progresses. Active focus trains the model to try new approaches when a task gets hard, ask for clarification and check in before important actions.
Meta also lists multi-step reasoning, tool use across long sessions and multimodal input such as screenshots, charts, d