Meta’s Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full has a narrower announced access path: Meta says it is available to Muse Code subscribers. Neither announcement establishes that the models have already replaced the engine behind Muse.
These access differences are the most important practical detail in Meta’s October 12 announcement. The company describes Wonder and Wonder Full as its most capable AI models yet for personal and professional agentic tasks, building on the work behind Muse Spark. Developer availability, subscriber access, and the planned consumer rollout are three separate parts of the release.
Meta also makes substantial claims about coding, domain expertise, memory, sustained execution, and safety. Those claims come from its own evaluations. They give developers reasons to investigate the models, but do not independently establish that Meta now leads every relevant category.
Wonder Is Available Now; the Muse Upgrade Is Still Ahead
The access map in Meta’s research announcement is explicit:
| Model or experience | Announced access | Status |
|---|---|---|
| Muse Wonder | Meta Model API | Available now |
| Muse Wonder | Muse Code | Available now |
| Muse Wonder Full | Muse Code subscribers | Available now for subscribers |
| Muse experiences powered by Wonder | Subscribers first | Coming later, without an announced date |
For developers building their own applications, Wonder has a clearly announced API route. These launch pages do not announce equivalent public API availability for Wonder Full. Access to Full inside Muse Code should not be read as permission to call Full from an arbitrary application.
Muse Code users have access to the broadly announced Wonder integration, with Full reserved for subscribers. The announcement does not spell out subscription prices, usage allowances, or whether particular subscription tiers carry different limits.
The consumer transition is still ahead. Meta says Wonder will “soon” power experiences in Muse, rolling out first to subscribers. Wonder will also power artifact generation and processes requiring more intelligence, with specialized models intended for broader Muse use further ahead.
Those plans do not establish a shipped replacement for Muse’s current engine, and Meta gives no date for that replacement. Today’s release can be judged on the announced API and Muse Code access; the future Muse experience has yet to arrive.
A Third-Party Listing Adds Pricing and Context Details
The official announcement explains the release’s capabilities more clearly than its operating costs. A separate OpenRouter model listing provides pricing and specifications for Wonder’s API offering, though these come from a different source.
OpenRouter lists Muse Wonder at $4 per million input tokens and $20 per million output tokens, with cached input reads at $0.20 per million tokens. It lists a 1,048,576-token context window and describes support for text, images, video, audio, and PDF documents as inputs, with text output.
These are OpenRouter’s listed details, not a quotation of Meta’s direct API pricing, Muse Code subscription terms, or specifications for Wonder Full.
For developers working with repositories or lengthy document collections, a large context window answers one question: how much material the service can accept. It does not establish how accurately the model will retrieve a particular detail, reconcile conflicting documents, or finish a long-running task.
Token rates also cover only part of an agent’s cost. A workflow that repeatedly reads files, revises a plan, or retries failed actions needs an end-to-end cost measurement. Given the launch’s emphasis on sustained work, that measurement is more useful than a comparison based solely on the price of one prompt.
Active Memory and Active Focus Target Different Problems
Meta organizes much of Wonder’s agentic pitch around two named capabilities: Active Memory and Active Focus.
Active Memory is the company’s term for carrying important context across conversations and retrieving it when it becomes relevant. The aim is to make an agent useful beyond the immediate exchange without requiring the user to restate every preference or earlier decision.
Cross-conversation recall is distinct from the context-window specification. A context window concerns information supplied within a request; recall concerns what persists between interactions and how the system decides to bring it back.
The launch text does not explain enough to equate that product-level memory promise with automatic persistence in every API integration. Developers should verify where memory is stored, what the application must provide, and which controls exist for inspecting, correcting, or deleting remembered information.
Active Focus addresses execution. Meta says Wonder can keep working toward a difficult goal, try alternative approaches, recover from errors, adapt to new information, and ask for clarification. The company also describes check-ins before important actions.
In coding, the intended benefit is sustained progress through requirements, code changes, testing, and revision. Meta says it trained both models on diverse, long-running tasks to support that process.
The two capabilities address different problems, but their names do not establish reliability. An agent that remembers the wrong constraint can repeat a mistake across sessions. Without an effective stopping rule, persistence can mean spending more time on an unsuitable approach. Evaluation needs to examine successful completion as well as behavior when a task becomes ambiguous or blocked.
Meta’s Benchmark Claims Need Their Testing Conditions
Meta says the Wonder family achieves state-of-the-art results across benchmarks for personal and professional agents, memory, and coding. The launch’s specific leadership claims are vendor-reported, not independently reproduced rankings.
The research post names several evaluations with different purposes. Harvey’s Legal Agent Benchmark, or LAB, tests work with source documents and detailed instructions to produce legal work products. Meta uses FocusBench to assess focus and SWE-Marathon to examine sustained execution on long-running software projects.
Answering a legal question, maintaining a goal over many steps, and revising a codebase require different capabilities. A strong result in one category cannot substitute for evidence in another.
The available source text does not expose the numerical results shown in the announcement’s charts, so there is no defensible basis here to reproduce a score table or calculate margins over competing models. Meta points readers to an evaluation report, but the launch prose alone is not enough to audit every comparison condition.
Developers assessing the results need to know the exact model versions, reasoning effort, available tools, agent harness, and execution budget. Those conditions help distinguish model capability from the contribution of the surrounding software and the resources allowed during testing.
Meta also presents a generalization comparison between Muse Spark 1.3 and Wonder Full. At different reasoning efforts, it compares performance on tasks targeted during post-training with performance on tasks it says were not targeted.
According to Meta, Wonder Full performs better on unfamiliar tasks than the Spark 1.3 trend would predict. The company attributes that improvement to larger, stronger base models whose post-training gains transfer more effectively.
This is a more substantive argument than a leaderboard position alone: it attempts to address whether training improvements extend beyond the tasks emphasized during development. The task selection and explanation still come from Meta. The result supports its case for broader capability without independently establishing generalization across the full range of work developers might assign.
Sources
- Meta’s October 12 announcementabout.fb.com
- Meta’s research announcementresearch.meta.ai
- OpenRouter model listingopenrouter.ai
