On October 12, 2026 (U.S. time), Meta released the AI model Muse Wonder and its higher-end version, Muse Wonder Full. Meta describes the two as its "most capable models to date" and says it built them for AI agents that handle personal errands and work on a user's behalf. An AI agent is an AI that takes instructions and carries out multiple steps on its own, such as researching or operating apps. TechCrunch reported that this is Meta's first "frontier model" (a model on par with the cutting edge of the industry).
Differences between Muse Wonder and Muse Wonder Full
The Muse Wonder series follows Muse Spark, released in April 2026. Muse Spark is the model that powers Muse, the AI agent Meta announced in September. According to Meta, the models were trained at its large data center Prometheus. According to TechCrunch, this data center is in Ohio.
The two models differ in performance and where they can be used.
- Muse Wonder: available in the Meta Model API for developers and the coding agent Muse Code
- Muse Wonder Full: a more capable version, available only on some paid Muse Code plans. Not offered through the API
According to Meta's developer documentation, Muse Wonder reads text, images, video, audio and PDFs, and responds in text. The amount it can handle at once (context) is about 1 million tokens. The API version is currently treated as a "preview" (a stage before the official release).
The model weights (the data that makes up the AI) have not been released. In Meta's documentation, Muse Wonder is classified as a model served on Meta's servers, while the open model that can run on your own computer is a separate line called Muse Glimmer.
Key features: Active Memory and Active Focus
Meta highlights two capabilities as defining features of Muse Wonder.
- Active Memory: remembers important information about the user across conversations and proactively recalls and uses what is relevant to the current task
- Active Focus: keeps working toward difficult goals by trying other approaches, overcoming obstacles and adapting to new information. It asks for clarification when needed and checks with the user before important actions
According to CNBC, Alexandr Wang, head of Meta's AI division, said: "In real-world tasks you always hit walls. Rather than stopping, we trained it to try another approach and to be able to check with the user at any time."
Meta says the models are also strong in legal, financial and medical knowledge, and in long-running programming work. Its official blog shows examples such as creating a 3D game from an image and planning a running schedule by operating a browser and multiple apps.
Benchmarks published by Meta
Meta published results comparing Muse Wonder with Anthropic's Claude Fable 5.1 and Claude Opus 5.5 and OpenAI's GPT-6 Astra. As a rule, the comparison uses each model's deepest thinking setting (max), in a table compiled by Meta. The table mixes figures from external evaluators such as Artificial Analysis, Vals AI, Surge and Datacurve, figures published by each company and figures Meta measured itself.
The main areas where the Muse Wonder series came out ahead are as follows (in the order Muse Wonder Full / Muse Wonder / Fable 5.1 / GPT-6 Astra / Opus 5.5).
- MemoryBench (ability to remember information within long conversations; Meta's own evaluation): 85.5 / 85.6 / 79.0 / 81.4 / 81.5
- Harvey LAB-AA (drafting practical legal documents): 13.1 / 10.0 / 6.4 / 8.6 / 4.2
- DeepSWE 1.1 (long-running software development): 79.0 / 78.8 / 67.4 / 74.1 / 74.2
- SWE-Marathon (even longer software development): 57.5 / 53.1 / 45.6 / 42.5 / 46.3
The SWE-Marathon figure for Muse Wonder Full is at the "xhigh" setting, one step below the maximum. Also, GPT-6 Astra was measured in Codex and Fable 5.1 in Claude Code, so the comparison is not under identical conditions.
On the other hand, there are also areas where other companies' models came out ahead.
- HealthBench Pro (answering medical questions): 62.3 / 60.5 / 62.1 / 65.3 / 65.6
- Terminal-Bench 4.0-AA (development work in the terminal): 57.1 / 52.0 / 52.0 / 59.1 / 59.6
- ProgramBench (building programs from scratch): 59.0 / 50.0 / 55.5 / 52.5 / 70.5
- OSWorld V2.1 (operating a computer screen; evaluation including partial credit): 78.4 / 72.6 / 80.7 / 75.8 / 81.8
- GDP.pdf (work based on PDF documents): 33.2 / 30.8 / 27.6 / 34.2 / 30.6
- Artificial Analysis Intelligence Index (overall index): 56 / 52 / 53 / 53 / 58
Of these, only Claude Opus 5.5 beat Muse Wonder Full on the overall index and ProgramBench, while Muse Wonder Full beat Fable 5.1 and GPT-6 Astra. On answering medical questions, Claude Opus 5.5 and GPT-6 Astra came out ahead. Results vary by area, and it did not outperform other companies in every field.
Where it is available and pricing
The model ID in the Meta Model API is muse-wonder-1.0. According to Meta's pricing page, prices per million tokens are as follows.
- Input: $4.00
- Cached input (when the same content is sent repeatedly): $0.20
- Output: $20.00
Muse Spark has Contributor mode, which is cheaper in exchange for Meta using the content you send to train its models, but it is not available with the Muse Wonder API.
Muse Code can also be used through monthly subscription plans. Muse Wonder is available on all plans starting from $5 per month, and Muse Wonder Full is available on Power Usage at $50 per month and Maximum Usage at $200 per month. On paid Muse Code plans, Contributor mode, which gives you more usage in exchange for Meta using what you send for training, is turned on by default and can be turned off in settings.
Muse Wonder will also soon become available in the AI agent Muse, rolling out first to paid-plan subscribers. CNBC reported that using it in the Muse app requires a $20 or $100 monthly plan.
Availability in Japan and caveats
Meta's announcement does not mention availability in Japan. The geographic use policy, part of the Meta Model API terms of use, states that the API is "not available in all countries" and lists countries and regions where it cannot be used, and Japan is not on that list. However, the same policy says it can be used only in countries and regions where Meta has enabled access, and it is not officially stated whether you can actually sign up and use it from Japan, so this needs to be checked. The AI agent Muse itself has been reported to have launched in the U.S. in September and later expanded to Canada, but availability in Japan has not been announced.
On safety, Meta says it has strengthened safeguards against misuse for chemical, biological, radiological and nuclear (CBRN) weapons and cyberattacks more than ever before. According to the model documentation, because it could not rule out that the model, without mitigations, could reach a high-risk level in the chemical and biological domains, Meta built in additional safety training and mechanisms to detect inappropriate use, and assessed the risk after mitigations as "medium or lower." It says a detailed safety report will be "published soon." The model documentation also states that reasoning involving space, such as 3D, remains a weakness, and that the model is not intended for use by people under 18.
CoAI's take
Muse Wonder is a model in which Meta puts front and center "the ability to remember" and "the ability to see things through," capabilities that matter when AI agents are used for real-world tasks. In Meta's own comparison, it outperforms other companies' top models in memory and long-running development work, while falling short of Claude Opus 5.5 and others in some areas such as medical answers, screen operation and the overall index, so its strengths are clear-cut.
With API pricing published at $4 for input and $20 for output, it has become easier to compare costs with other companies' top models. However, it is currently a preview, and the official terms of availability in Japan and the timing of its arrival in the Muse app have not been disclosed. We will need to wait for further news on whether Muse can be used in Japan, including the dedicated device Muse Charm announced at Meta Connect in September.
Sources
- Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents (Meta)
- Muse Wonder: A Frontier Model For Personal & Professional AI Agents (Meta AI Research)
- Muse Wonder Model Card (Meta AI Research)
- Models (Meta Model API)
- Pricing and rate limits (Meta Model API)
- Muse Code (Meta)
- Meta Model API Model Designation and Geographic Use Policy (Meta)
- Meta releases its first frontier model, Muse Wonder (TechCrunch)
- Meta debuts Muse Wonder AI model in latest test of users' willingness to pay (CNBC)
- Introducing Muse: The World’s First Personal AI Agent Built for Everyone (Meta)
- Meta’s Muse AI Agent Is Now Available in Canada (iPhone in Canada)
