On 12 October, Meta introduced Muse Wonder and Muse Wonder Full. Wonder is available through the Meta Model API and Muse Code; Full is available to Muse Code subscribers. Deployment in the consumer Muse experience is a subsequent announced step. That distinction matters: launching a model does not mean every application or account already uses it.

The notable shift is the emphasis on sustained work. A single answer can appear convincing while a project requires decisions to survive, results to be checked and unfinished work to be resumed. Our editorial reading of the announcement is this move from an answer toward a continuing process.

Memory and focus

Meta's research post calls the capabilities Active Memory and Active Focus. The first retrieves relevant context from earlier conversations. The second describes continuing toward a goal through alternative approaches, recovery from errors and clarification. Meta identifies FocusBench and SWE-Marathon as evaluations of sustained execution. These are the developer's descriptions and evaluations, not our own independent behavioral tests.

An application project illustrates the distinction. Useful memory should recover the agreed requirements; useful focus should complete a verified change. Remembering an incorrect requirement and pursuing it persistently would produce a consistent but wrong result. Those capabilities therefore need to be assessed alongside the quality of the checks around them.

What the documentation establishes

The model card lists text, image, video and audio inputs, text output and a one-million-token context. It also warns about inaccurate responses, errors in multi-step tasks and weaknesses in visual and spatial reasoning. Tools and the agent's operating environment can affect performance. A large context window consequently does not guarantee that every detail will be used correctly.

Vercel's catalog provides a separate check that Wonder is offered through AI Gateway. Its listed prices are $4 per million input tokens and $20 per million output tokens. This establishes a specific access channel and a published tariff; it does not establish superior task performance. We do not treat the catalog's speed figures or vendor comparisons as universal results.

There is also a practical difference between the model's input capacity and an application's durable memory. Our interpretation is that a deployment should show which information is retained, which is retrieved for a particular job and how a user corrects it. Sending a long document into a context window is not, by itself, evidence that a project will be remembered accurately next week.

How to assess usefulness

Our editorial assessment is that a small real workflow with predefined criteria offers a useful test. An agent could change an application, preserve the reason for the decision and recover that same reason in a later session. A separate check should establish whether it stops before an action that changes access or publishes information. Memory is valuable when its contents can be inspected and corrected.

This news establishes the availability of new models and the direction of their design. SIExpo has not run its own comparative evaluation and does not claim that Wonder replaces professional judgment or guarantees error-free autonomous work.

Sources checked on 13 October 2026. Availability and prices reflect the time of verification. The cover is an AI illustration, not a photograph of an actual event.

Editorial verification (UTC): 2026-10-13T06:10:05.979402+00:00