What is new

Meta's Superintelligence Lab released two new models on Monday: Muse Wonder and Muse Wonder Full. Meta describes these as its first frontier model, a step up from the earlier Muse Spark, which was built to be lightweight and fast rather than maximally capable. The Wonder models are larger and, according to Meta, perform better on complex tasks such as coding, with benchmark scores the company says are comparable to two other recently released frontier systems, Fable 5.1 and GPT-6 Astra.

A frontier model, in industry usage, generally refers to the most capable class of AI system a lab can currently build, typically the largest and most expensive to train and run, and the one used as a benchmark for what AI can currently do. Labs often also release smaller, cheaper, faster models alongside these for everyday use. Muse Spark appears to have served that role for Meta before Wonder arrived.

How it works

The headline feature Meta is promoting is something it calls "active memory": the ability of the model to retain and synthesize information a user gives it over long stretches of time. In general terms, this relates to a broader goal in AI research sometimes called continual learning, where a system is meant to keep useful information from past interactions rather than treating every conversation as if it started from zero. Standard large language models typically only "remember" what fits inside a session's context window; anything outside that is usually lost once the conversation ends, unless a separate memory system stores and retrieves it.

To support this, Meta pointed to a benchmark called MemoryBench, which it says tests continual learning and memory construction frameworks. According to the company's own testing, Muse Wonder scores higher on MemoryBench than Fable, Opus, and Astra. Research scientist Kavya Kopparapu said in a launch video that the team put significant effort into making sure the model can remember what matters to a user over days, weeks, months, and beyond.

Both Wonder and Wonder Full are also built to power Meta's Muse agent, the company's assistant product designed to carry out tasks over an extended period rather than answer single questions. An AI agent, broadly speaking, is a system that can take a goal, break it into steps, use tools, and continue working over time rather than just producing one response, so better memory retention is directly relevant to how well such an agent can track a long-running task or a user's preferences.

Why it matters

This release matters because it signals that Meta is now competing directly in the frontier-model tier alongside OpenAI and Anthropic, something it had largely avoided since restructuring its AI division. The pricing also places Wonder squarely in that competitive bracket rather than as a budget option.

  • Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API
  • Meta says that price is roughly equivalent to GPT-6 Sol or Opus 5.5
  • Muse Wonder Full will only be available to Muse Code subscribers and will not be offered through the API
  • Both models power the Muse agent product
  • Muse Wonder is reported to outscore Fable, Opus, and Astra on the MemoryBench benchmark
  • The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year

Tokens are the small chunks of text (roughly pieces of words) that models process and generate; usage-based pricing per million tokens is the standard way providers charge for API access, with output tokens usually costing more than input tokens because generating text is more computationally expensive than reading it. At $4 input and $20 output per million tokens, Wonder sits in a premium pricing tier typical of top-end frontier models, rather than the cheaper rates associated with lightweight models like Muse Spark.

The release also comes roughly a year after Scale AI founder Alexandr Wang restructured Meta's AI efforts, in August of the previous year. Since that reorganization, Meta's public output had consisted of the lightweight Muse Spark model and the Muse agent, without a state-of-the-art flagship release of the kind OpenAI and Anthropic have become known for. Wonder and Wonder Full are presented as the first result of that restructuring reaching frontier-level performance, which Meta frames in its blog post as progress toward what it calls personal superintelligence.

Limits and open questions

The benchmark comparisons described here come from Meta's own testing and model card, not from independent third-party evaluation, so practitioners should treat the head-to-head claims against Fable, GPT-6 Astra, Opus, and GPT-6 Sol as Meta's self-reported figures rather than confirmed external results. The announcement does not specify how MemoryBench is constructed, what tasks it contains, or how large the performance gap is between Wonder and its rivals beyond stating that Wonder scores higher. It also does not detail what makes the Wonder Full variant different from the standard Wonder model beyond its restricted Muse Code-only availability, nor does it give concrete coding benchmark numbers, only a general claim of better performance on complex tasks like coding.

How to try it

Muse Wonder is accessible through the Meta Model API at the pricing listed above. Muse Wonder Full is restricted to subscribers of Muse Code and is not offered through the API, meaning developers who want to integrate Wonder Full directly into their own applications via API calls currently cannot do so.

Sources