An agent may write a function in seconds, then lose its way when it has to edit multiple files, fix a failing test and revisit a decision made hours earlier. Meta is targeting that second problem with Muse Wonder and Muse Wonder Full, introduced on October 12 for coding and work that requires continuity.
The company describes Active Memory as a way to retrieve relevant context and Active Focus as the ability to keep working through obstacles. These are claimed capabilities, not a promise that an agent will complete any project. The practical aim is to preserve the goal throughout a task rather than produce one strong standalone answer.
What developers can use today
Availability matters because similar names conceal different access paths.
| Variant | Confirmed launch access |
|---|---|
| Muse Wonder | Meta Model API and Muse Code |
| Muse Wonder Full | Muse Code Power Usage and Maximum Usage plans |
Wonder is in preview. Its API model ID is muse-wonder-1.0, with a 1,048,576-token context window. Choosing maximum reasoning effort does not select Full.
Pricing per million tokens is $4 for input, $0.20 for cached input and $20 for output. As a worked example, 100,000 uncached input tokens and 20,000 output tokens cost $0.80 before additional services. Retries, tool use and corrections can involve multiple calls. Token pricing alone does not tell you what completing the task will cost.
Memory needs operating rules
Remembering constraints and earlier decisions could spare a team some rework. MnzAI Labs' view is that the useful test involves changing context: does the agent drop an old instruction when corrected? Does it retrieve the right decision, or persist with a plan the team has abandoned?
There is also a concrete data choice. Meta says content in Standard mode is not used to train its models. Muse Code subscriptions default to Contributor mode, which permits training use in exchange for more usage and can be disabled in settings. Wonder does not offer Contributor mode on the API. Before connecting a company repository, that setting deserves as much attention as tool permissions.
An agent that remembers more also needs to recognize when an earlier decision no longer applies.
A leaderboard is not a deployment forecast
The model card describes coding and professional-work evaluations, but comparisons do not always share a common execution environment. SWE-Marathon, for instance, includes different agent products and configurations. That makes the scores useful leads, with limited power to predict results in your own repository.
Meta also plans to bring Wonder to its Muse assistant, beginning with subscribers. That future rollout is separate from today's developer access.
The next useful experiment is a representative task with acceptance criteria and a spending limit set before the run. Measure the outcome, human interventions and recovery failures. If the model sustains progress with less supervision while respecting data boundaries, the team has an operational improvement to discuss beyond its leaderboard position.
Sources: Meta research, announcement and rollout, developer guide, model availability and data modes, pricing, API overview and model card and evaluation methodology.