TL;DR: Muse Wonder is a large frontier model designed for complex tasks, coding, and active memory (a system that continuously remembers user information). It is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, while Muse Wonder Full is limited to Muse Code members.
Meta has launched Muse Wonder and Muse Wonder Full, the company’s first frontier models, positioning them to handle complex tasks such as coding more effectively and to remember information provided by users through the concept of “active memory” (a system that continuously remembers user information). They are suited to tasks that continue over several days or weeks.
Muse Wonder is available through the Meta Model API (a channel for programs to call the model) at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is available only to Muse Code members and is not offered through the API.
An Overview of Muse Wonder
Muse Wonder and Muse Wonder Full are Meta’s first frontier model series. They are larger than Muse Spark and place greater emphasis on complex tasks such as coding. Meta’s test results indicate that their benchmark scores (standardized tests for comparing capabilities) are close to those of Fable 5.1 and GPT-6 Astra.
Both models will also power the Muse agent. Muse Wonder is accessible through the Meta Model API, while Muse Wonder Full is limited to Muse Code members.
When an AI Assistant Needs to Remember Us Longer Than a Single Conversation
If we need to talk with AI about the same subject continuously for days, weeks, or months, remembering important information and connecting it to past context is the core of a personal assistant.
Muse Wonder emphasizes this through the concept of “active memory,” which focuses on continuously remembering and synthesizing information from users rather than simply answering questions in a single conversation. Meta states that Muse Wonder scored higher on MemoryBench than Fable, Opus, and Astra, reflecting the model’s strength in long-term information retrieval.
Muse Wonder sits above Muse Spark in terms of size and intended use. Wonder Full is available only to Muse Code members.
| Factor | Muse Spark | Muse Wonder | Muse Wonder Full |
|---|---|---|---|
| Product positioning | Lightweight model | Large model for complex tasks | Large model for complex tasks |
| Capabilities | Focuses on low latency | Improved benchmark results and coding performance | Improved benchmark results and coding performance |
| Use cases | Tasks requiring fast responses | Complex tasks and the Muse agent | Complex tasks and the Muse agent |
| Availability | Not specified in the source | Meta Model API | Muse Code members only; not available through the API |
Overall, Muse Spark is like the younger sibling that emphasizes flexibility, while Muse Wonder is the big sibling for tasks requiring greater capabilities. If it were me, I would choose based on the nature of the work rather than simply looking at the model name.
If you want to access it through an API, the choice is Muse Wonder. Wonder Full is better suited to people who are already part of the Muse Code membership system.
From Muse Spark to Frontier Models
Muse Spark is positioned as a lightweight model focused on fast responses, or low latency (minimal delay). Muse Wonder and Wonder Full move up to become large models for complex tasks, particularly coding and working with the Muse agent.
| Factor | Muse Spark | Muse Wonder | Muse Wonder Full |
|---|---|---|---|
| Overall size | Lightweight model | Larger than Muse Spark | Larger than Muse Spark |
| Capabilities | Focuses on low latency | More capable | More capable |
| Coding | Not specified | Improved results on complex tasks | Improved results on complex tasks |
| Test scores | Not specified | Close to Fable 5.1 and GPT-6 Astra | Close to Fable 5.1 and GPT-6 Astra |
| Role in AI strategy | Lightweight model | Frontier model that powers the Muse agent | Frontier model that powers the Muse agent |
Comparing Muse Spark with Muse Wonder is like choosing the right tool for everyday work: Spark is suited to tasks that require speed, while Wonder is better for problems that require thinking through and connecting multiple steps.
Wonder is available through the Meta Model API, while Wonder Full is limited to Muse Code members and is not available through the API.
Capabilities That Could Change How We Use AI in Real Life
-
Complex tasks: Wonder is a model with higher benchmark scores and improved performance on complex tasks, making it suitable for problems that require multi-step reasoning.
-
Coding: Meta states that Wonder delivers better coding results, making it suitable for reading code, troubleshooting, and handling more complex development work.
-
Continuous memory: The “active memory” feature helps the model continuously remember important information about users over several days, weeks, or months.
-
Information synthesis: When users provide information, Wonder can remember, summarize, and connect that information. Its MemoryBench test results were higher than those of Fable, Opus, and Astra.
How Does Muse Wonder Compare with Fable, Astra, and Opus?
Overall, according to the news, Wonder’s benchmark scores are close to those of Fable 5.1 and GPT-6 Astra, based on Meta’s test results. Its MemoryBench score is higher than those of Fable, Opus, and Astra. The pricing is only a comparison of API usage costs, not a performance test.
| Factor | Muse Wonder | Fable 5.1 | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|---|
| Benchmark score (Meta) | Close to Fable 5.1 and GPT-6 Astra | Used as a comparison benchmark | Used as a comparison benchmark | Not specified |
| MemoryBench performance (Meta) | Higher than Fable, Opus, and Astra | Lower than Wonder | Lower than Wonder | Lower than Wonder |
| Price | $4 input / $20 output; close to the prices of GPT-6 Sol and Opus 5.5 | Not specified | Not specified | No price figures |
Key Strengths and Limitations to Know Before Using It
Muse Wonder is a frontier model with benchmark scores close to those of Fable 5.1 and GPT-6 Astra, along with MemoryBench test results higher than those of Fable, Opus, and Astra, making it suitable for complex tasks such as coding.
I recommend viewing Muse Wonder as an option for tasks that require continuous context and coding, rather than expecting it to be suitable for every task, because Wonder Full is still limited to Muse Code members and the reported benchmark results come from Meta itself.