Tech • AI • Robotics • Game
Meta's New Muse Wonder Is Huge!
9.2/10TL;DR
Meta has launched Muse Wonder, a new agent-focused model family that it says can compete with top systems from OpenAI, Anthropic, and Google, with rollout starting through its API, Muse Code, and later the Muse app.
KEY POINTS
A new model family for agents
Muse Wonder includes two versions: Muse Wonder and Muse Wonder Full, with Full positioned as the larger and more capable model. Meta built both specifically for AI agents that can carry out practical multi-step tasks rather than just answer prompts. Target use cases include insurance claims, lease review, personal finance, health-related assistance, and app building.
Two core features stand out
The release centers on active memory and active focus. Active memory is designed to retain relevant user context across conversations so people do not need to repeatedly restate preferences or background information. Active focus is meant to keep the agent working through difficult tasks, trying alternate approaches, handling errors, and asking clarifying questions instead of stopping midway or taking unwanted actions.
Consumer-style demos emphasize autonomy
In one example, the model built a personalized running plan in a browser by accessing a calendar and a map, then plotting routes based on a user’s schedule and preferences. In another, it created and ordered a custom birthday T-shirt for a cat named Coco, pulling a photo from Instagram, editing the design, and completing the purchase flow. These examples were presented as evidence that the system can navigate websites and execute real-world actions.
Coding and creation are part of the pitch
Meta also showcased a browser-created game called Muse Kitchen, where players move around a kitchen, pick up ingredients, chop them, and plate them. The game was simple, but it was used to demonstrate that the model can generate functioning interactive software with usable graphics and mechanics. That aligns with Meta’s push to market the model for agentic coding and app development.
Personal context is central to the Muse app
Inside the Muse app, one demonstration showed the model responding to travel-related symptoms by checking a user’s calendar, messages, and email. It inferred that the person was traveling and suggested altitude sickness as a likely explanation. The example was intended to show how long-term memory and cross-app context could improve personalized assistance.
Administrative work is a major use case
Another demo focused on an insurance claim after an urgent care visit. After reviewing uploaded images, billing details, and an insurance card, the system identified a possible payout error, proposed calling the insurer, and then shifted to drafting an email when the user said there was no time for a phone call. The scenario highlighted a category of work that many users find especially frustrating and time-consuming.
Complex planning tasks are another target
Meta also showed the model preparing a wedding seating chart based on detailed social constraints and conflicting preferences. The example underscored the company’s broader argument that agent systems can help with messy real-world planning, where many variables need to be balanced at once.
Benchmark claims suggest Meta has narrowed the gap
On Meta’s own Personal Super Intelligence Index, Muse Wonder Full scored 59.1, ahead of Claude Opus 5.5 and GPT-6 Astra in the company’s comparison. On SWE Marathon, a long-running coding benchmark, Meta reported 57.5 for Muse Wonder Full versus 46.3 for Opus 5.5. However, an external ranking from Artificial Analysis reportedly places Muse Wonder Full at 56 on its intelligence index, in the top tier but still behind Claude Opus 5.5.
Rollout is beginning in stages
Muse Wonder is now available to developers through Meta’s API and Muse Code, while Muse Wonder Full is being offered to Muse Code subscribers. Meta says the new model will also power the Muse app, with access expected to reach paid subscribers first and later expand more broadly, including eventually to the free plan.
CONCLUSION
Muse Wonder appears to be Meta’s strongest AI model release to date and a serious attempt to compete at the frontier of agent systems. If its performance and autonomy hold up in wider use, it could strengthen Meta’s position in consumer AI and increase pressure on rivals to improve their own agents.
Ask a question