On October 12, Meta announced two AI models,
Muse Wonder and Muse Wonder Full.
They feature Active Memory, which carries important information
across conversations, and Active Focus,
which keeps working toward a goal even when a task gets difficult.
Muse Wonder is now available through the API and the Muse Code development tool,
and Meta also plans to bring it to Muse, its personal AI agent.
What Was Announced: Two Models for Development and Real-World Work
Meta positions the Muse Wonder series as its most capable models to date,
built for personal and professional agentic tasks and coding.
Examples include navigating an insurance claim, understanding a lease,
and building games or business apps.
Availability differs by model.
Muse Wonder is available through the Meta Model API and Muse Code.
Muse Wonder Full is included with Muse Code subscriptions,
and the developer page lists the Power Usage and Maximum Usage plans.
Muse Wonder API pricing per million tokens is
US$4 for input, US$0.20 for cached input, and US$20 for output.
The Muse Code plans that include Muse Wonder Full
are US$50 per month for Power Usage
and US$200 per month for Maximum Usage.
What Changes: Carrying Over Past Conversations and Recovering Mid-Task
Active Memory is the ability to carry information that matters to the user
across conversations and bring back details relevant to the current task.
Meta explains that it built this capability
to handle real work that unfolds over time.
Active Focus is described as the ability to keep working toward a goal
when a task gets hard, try other approaches, and recover from errors.
Meta says it is designed to ask for clarification when needed
and to check in before carrying out anything important.
A 1-Million-Token Context and Support for Multi-Step Work
The official developer page lists a context length
of 1 million tokens for both models.
Use cases include making and verifying changes
across multiple files in large repositories,
and operating browsers, desktop, and mobile apps.
The page also explains that for multi-step tasks running for hours,
the model plans the work, delegates parts to sub-agents,
and calls tools in parallel.
These are capabilities described by Meta
and do not guarantee the same results for every task.
Available for Developers Now, Coming to Muse Later
Muse Wonder will also power Muse, Meta’s personal AI agent, in the future.
Meta says it will roll out first to subscribers.
Availability through the API and Muse Code
and the planned rollout to Muse are presented as separate stages.
With agents that work across messages, calendars, and documents in mind,
Meta says it trained the models to avoid sharing private information
without permission.
A full report including additional safety evaluations
is also expected to be published.
LC Signal’s Perspective: Using Conversational Memory for Ongoing Work
The defining feature of this release is the combination
of carrying important information across conversations
and recovering by trying other approaches mid-task.
If conditions shared earlier can be used in later work,
and the approach can change when progress stalls,
the range of work that can be delegated on an ongoing basis
seems likely to grow.
The models are already available to developers,
while the rollout to Muse for individuals is still to come.
We will watch how the memory and task-continuation capabilities work
both in coding and in everyday use involving calendars and documents.
Key Sources
-
Meta official announcement:
Introducing Muse Wonder and Muse Wonder Full:
Frontier Models for Agents
-
Meta official developer model page:
Muse Wonder
This article was prepared with AI assistance using publicly available information
as of October 13, 2026.
Announcement dates follow the dates shown on the official pages.
Current availability is distinguished from the planned rollout to Muse.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.