Meta Releases Muse Wonder, Its First Frontier AI Model

Meta Muse Wonder: What to Know About Meta's New Frontier AI Model

Meta has released Muse Wonder and Muse Wonder Full, its first frontier-model offerings since reorganizing its AI operations. The company is highlighting the models' performance on MemoryBench, a benchmark focused on how AI systems retain and build on information over time.

The launch marks a step in Meta's effort to develop models for more persistent, personalized interactions. Meta says the models can use information gathered across days, weeks, and months to make responses more useful. That is a company claim about their intended capabilities, not a guarantee that every interaction will preserve or apply information reliably.

Meta Introduces Muse Wonder and Muse Wonder Full

Meta describes Muse Wonder as a larger and more capable model than Muse Spark, its earlier model. The company also points to improvements on complex tasks, including coding. Muse Wonder Full is the more capable version in the new family, but Meta is limiting access to Muse Code subscribers.

The release is significant because it is Meta's first frontier-model launch since its AI reorganization and the formation of Meta Superintelligence Labs. Rather than presenting the models only as general-purpose chat tools, Meta is emphasizing their potential to remember useful context and support tasks that continue across sessions.

What Meta's MemoryBench Results and Model Claims Show

Meta says Muse Wonder scored higher on MemoryBench than Fable, Opus, and Astra. The benchmark is designed to assess continual learning and memory construction - whether a model can identify, retain, and use information as it accumulates.

That focus reflects a broader challenge for AI assistants. A system may produce a strong answer in a single conversation but be less useful if it cannot keep track of relevant preferences, facts, or ongoing work. Meta's stated goal is for Muse Wonder to make longer-running interactions more coherent by retaining helpful information over extended periods.

The reported result should be read in context: a benchmark score measures performance on a particular evaluation, not the full quality of a model in everyday use. Meta has not provided scores in the reported launch details, so readers cannot assess the size of the gap or compare the models' performance across other tasks from this result alone. Practical usefulness will depend on how well the models handle real tasks, including when information changes or is no longer relevant.

Muse Wonder Pricing, Access, and the Planned Muse Agent

Meta lists Muse Wonder at $4 per million input tokens and $20 per million output tokens. Input tokens cover the text and other supported content sent to the model; output tokens are the model's generated response. The pricing gives developers a way to estimate usage costs, though actual bills depend on how much information their applications send and how much the model produces.

Muse Wonder Full is available only to Muse Code subscribers, according to the launch details. Meta has not announced the same per-token pricing for that version in the information reported so far.

Meta also plans to use the models to power the Muse agent. That points to a role beyond answering isolated prompts: an agent can use a model to help carry out tasks through multiple steps. The planned integration makes memory especially relevant, since an agent may need to draw on prior context while working toward a goal. The announcement does not, by itself, establish what tasks the Muse agent will support or how broadly it will be available.

Muse Wonder gives Meta a new frontier-model offering, with long-term memory as its central pitch. Its MemoryBench result is an early signal of how Meta wants the models judged; pricing, subscriber access, and the planned Muse agent will shape how developers and users can put them to work. Explore our coverage of Meta's AI models and the latest developments in frontier AI.