Skip to content

Meta Launches Muse Wonder: Stronger Models, Different Access Paths

AI

TL;DR

Muse Wonder has API access, while the stronger Full version initially serves Muse Code subscribers. Better long-task memory and independent results create reasons to consider it, but teams must distinguish accessible models, memory correction, and results inside a bundled product.

Meta Launches Muse Wonder: Stronger Models, Different Access Paths

Meta has introduced Muse Wonder and Muse Wonder Full, but initially offers API access only to the former. That leaves a concrete comparison limit: whether Full’s performance inside a coding product can justify choosing it for a custom agent remains unconfirmed until developers can run the same model in their own workflow.Availability

The official release date is 2026-10-12, without a time or timezone. TechCrunch timestamp: 2026-10-12T10:56:00-07:00. This article uses a 09:07 Taipei cutoff on October 13 and explicitly draws on the previous day’s release within the 48-hour window. A report crossing midnight does not establish a same-day event.Release Reporting

Wonder provides a 1M-token context window. Its API costs $4 per million input tokens and $20 per million output tokens. Full has no API initially and is available through Muse Code’s $50 and $200 monthly plans. Subscription and usage pricing serve different access arrangements; they cannot be directly converted into the cost of an identical workload.API and subscriptions

Memory for long tasks must allow corrections

Meta’s Active Memory carries context across conversations and retrieves relevant details. Active Focus tries alternative approaches when blocked and asks users for clarification when needed. The company demonstrated a playable 3D kitchen game built with Wonder. That illustrates an interactive generated result, not evidence that ordinary users can build arbitrary software to the same standard.Mechanisms and demonstration

This direction looks more useful to me for work that undergoes repeated revision. Suppose someone designing a kitchen specifies a material, then changes the budget. The agent needs to withdraw affected decisions as well as remember earlier preferences. Retaining information may reduce repeated explanations, but an obsolete constraint governing later revisions could amplify a misunderstanding. This is a hypothetical product scenario, not a measured result of the demonstration.

I would expose the constraints currently in use and let the user correct them directly. A long task’s goal can change; refocusing a model does not by itself guarantee that it follows the latest requirements. For a product team, inspectable memory and consistent updates after a correction are closer to that need than permanently retaining an entire conversation.

The stronger Full arrives through a product first

At maximum reasoning effort, Artificial Analysis scored Wonder and Full at 52 and 56 on Intelligence Index v4.3, versus 48 for their predecessor Muse Spark 1.3. Wonder cost $4.52 per index task versus $1.60 for that predecessor. Full has no per-token price or comparable per-task cost. These results are not quotes for a company’s work and do not establish equivalent benefits for Traditional Chinese tasks.Independent testing

The model card notes that SWE-Marathon uses different agent tools across models, not controlled identical conditions. Meta does not rule out Full reaching the high-risk threshold for chemical and biological threats without safeguards; with safeguards, it rates risk moderate or lower. That is the company’s assessment, and a complete safety report is still pending.Evaluation conditions and limits Report status

I would therefore treat this release as two product choices. Teams willing to work in Muse Code can assess the combined capabilities of Full and its ready-made tools. Teams embedding an agent in their own service currently have Wonder. Progress in the first arrangement is not automatically a gain in the second. Nor should the highest-ranking variant become a product promise that the available access route cannot support.

Full’s current availability through a bundled product makes the interface part of model competition. My inference is that teams retaining an existing workflow will care more about whether access opens and how much capability survives outside the ready-made product. Wonder’s API and Full’s subscription entry point are now identifiable. The reviewed sources do not directly compare their completion quality with identical tools and identical requirement changes.

The cover reuses this site’s Muse Code brand illustration, not a Wonder interface or launch photograph. DNS resolution failures prevented downloading the event image.

Sources:

More on this topic

Random article