JOURNAL / 2026.10.13
Meta launches Muse Wonder and Wonder Full for agents that remember and sustain work
The new models expand options for coding and extended tasks. The distinction between evaluated memory, available context and actual access determines how much of the announcement can translate into an application.
Meta introduced Muse Wonder and Muse Wonder Full on October 12. Wonder arrives on Meta Model API and Muse Code; Full is offered to Muse Code subscribers. Integration into the Muse assistant is announced for later, starting with subscribers. This is a release developers can build with and evaluate now, although the two variants arrive through different routes.
The effort deserves attention for the kind of work it aims to sustain. In the research announcement, Meta calls retaining relevant information across conversations and retrieving it when needed Active Memory. Active Focus describes persisting through a blocker, changing strategy and adapting to new information. A refactor spanning several files needs both: preserving a constraint agreed at the outset and recovering when tests contradict the plan.
Compared with Muse Spark 1.3, Meta presents larger, stronger base models. Its comparison examines whether increasing reasoning effort also improves tasks that were not targeted during post-training. The company observes an advantage for Wonder Full over Spark's trend. This is a relevant hypothesis about generalization, supported here by the manufacturer's own evaluation; it cannot predict every combination of tools and requirements in someone else's project.
The model card contains a particularly useful distinction. MemoryBench is an internal evaluation of conversations spanning 64,000 to 100,000 tokens, with distractors and information that conflicts or gets updated; a model judge scores the answers. That tests retrieval of changing details within the evaluated material. It does not by itself demonstrate reliable memory over months, effective deletion of a preference or separation between users.
The same card describes SWE-Marathon as substantial software projects graded through executable tests. Comparisons retain different agent environments, and Full uses xhigh reasoning effort. They provide evidence about systems under those conditions, rather than an isolated model comparison or a guarantee for any repository. We can take the advance seriously without turning a ranking into a promise of delivery.
For developers integrating the model, the access documentation resolves an important confusion: selecting Wonder's maximum reasoning effort on the API does not turn it into Wonder Full. Full is reserved for Muse Code's Power Usage and Maximum Usage plans. Wonder accepts the same request formats as Spark, making it easier to try in an existing integration. The overview identifies it as a preview and documents a context window of roughly one million tokens. That input capacity does not amount to a permanent memory store either.
The published rates are four dollars per million input tokens and twenty per million output tokens, with discounted cached input. An extended task accumulates calls, context and possible retries: the relevant cost is reaching an acceptable result. Muse Code subscriptions also enable Contributor mode by default, permitting content to be used to train future models; it can be turned off. Wonder's API offers only Standard mode, whose content is not used for training according to that documentation. Choosing an access route also requires reviewing that agreement about data.
My assessment is that the launch expands useful competition: another option for extended work allows comparing behavior on the same project alongside published scores. I would start with a real, bounded change, introduce a requirements correction halfway through and check whether it reaches the relevant files and tests. I would retain errors, necessary interventions and the full cost. Useful memory needs to correct itself, as well as repeat what it first heard.
I would also keep visible what information the application stores, where it came from and how it is replaced. An agent that fluently retrieves an obsolete preference can be wrong with greater confidence. Wonder deserves that evaluation because of its ambition and concrete access; our trust should grow with evidence from work we can inspect.
Primary sources accessed October 13, 2026. We have not run the models or checked their results experimentally.