Meta's Muse Wonder Tests the Promise of

Meta introduced Muse Wonder and Muse Wonder Full on October 12 with a larger ambition than producing better answers to prompts. The company wants models that can remember relevant details across conversations and keep working when a task becomes difficult. Muse Wonder is available through Meta's Model API and Muse Code; Wonder Full is available to Muse Code subscribers. Meta plans to bring Wonder into its Muse personal agent, starting with subscribers. For now, the model release and the proposed agent experience occupy different stages of the rollout.

That sequence puts a useful question ahead of the leaderboard: can an AI model perform a demanding task repeatedly, over time, while keeping track of what the person asked it to do? A good answer in one session is helpful. An agent that carries an instruction into the next session has a longer chain of decisions to get right.

What Meta is Trying to Improve

Meta calls the two capabilities Active Memory and Active Focus. In the company's description, memory carries important context across conversations and retrieves it when a later task calls for it. Focus means trying another approach after a setback, adapting to new information, asking for clarification and checking before important actions. Those are descriptions of how Meta intends Wonder to behave, rather than separate assurances that every agent built with it will do so.

The product gives those terms a purpose. Muse arrived last month with an ambition to work across apps, draw on personal information and pursue tasks in the background. If an assistant merely answers a question, a forgotten preference may be annoying. If it continues a task across apps, the same lapse could change what it does next. Conversely, remembering everything indiscriminately would hardly be an improvement. An assistant must retrieve the detail relevant to the current instruction and recognize when the person needs to make the decision.

Meta says its evaluations cover personal and professional agents, memory and coding. Harvey's Legal Agent Benchmark asks models to work with documents and detailed instructions to produce legal work products. For sustained work, Meta names FocusBench and SWE-Marathon, the latter involving long-running software projects. A coding task makes the appeal of focus concrete: the assistant must keep requirements in view while making changes, testing them and revising its work. A system that stops at the first obstacle cannot deliver the project, however polished its opening suggestion.

Meta also argues that Wonder Full handles unfamiliar tasks better than its earlier Muse Spark 1.3 model would lead one to expect. Its comparison relates performance on tasks targeted during training to a separate set of unfamiliar tasks at different levels of reasoning effort. That is a sensible question for a personal agent, which will encounter combinations of goals and constraints its designers did not enumerate. The result remains Meta's evaluation of models, not a measure of how dependably Muse will handle a subscriber's ongoing work.

The practical test has several parts. The agent must recall the right information, use it under the current instruction, recover from a failed step and stop for approval when an action calls for it. Meta's stated design addresses those parts. Readers deciding whether to delegate work should look for how Muse handles corrections, changed instructions and actions requiring consent when Wonder reaches the product. That is a more useful test than asking whether a model once completed a difficult example.

There is also a safety question inside the product promise. A model that keeps trying after a setback is useful when the goal remains valid; checking before an important action gives the person a chance to say the goal has changed. Meta says it has trained Wonder to refuse dangerous requests. It has not disclosed whether agents powered by Wonder misbehaved in testing or what measures addressed that issue specifically. Those questions become more consequential when an assistant can continue working rather than wait for another prompt.

A Comeback with a Different Route to Users

Wonder follows a period in which Meta's earlier model strategy lost momentum. Llama's downloadable weights let researchers and startups run models on their own hardware, and the series gained broad adoption. Later Llama releases disappointed as Anthropic, OpenAI and Google introduced more capable models. Mark Zuckerberg reorganized Meta's AI effort last year, putting Alexandr Wang in charge of Meta Superintelligence Labs. Meta says it began a ground-up overhaul 14 months ago; Muse Spark appeared in April, followed now by Wonder. This history explains the pressure on the launch. Meta is asking users to judge its newest models after a period when distributing models widely did not keep it at the front of the capability race.

The Llama comparison also shows what has changed for people outside Meta. Llama's open weights offered a way to download a model and run it on one's own hardware. Wonder's launch offers an API or access through Muse Code instead. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, while Wonder Full is listed for Muse Code subscribers on $50- and $200-per-month tiers. A developer weighing the two approaches is deciding between direct control of a downloadable model and access to these newer models through Meta's services. That comparison concerns distribution; it does not imply Llama and Wonder have the same capabilities.

Meta has also released the smaller open-weight Muse Glimmer and says it intends to release larger open-weight models. Wonder's initial access terms therefore do not settle Meta's long-term position on open weights. They do determine who can use this release now and under what arrangement. For a researcher interested in running a model independently, an API offers a different opportunity from a download. For someone who wants an assistant inside Muse, the relevant issue is whether the planned subscriber rollout turns the model's stated abilities into dependable behavior.

Wonder gives Meta a model to put behind that promise. Its next test will take place across ordinary interruptions: a changed requirement, a failed attempt, a detail from an earlier conversation and a decision the assistant must leave to its user.