Meta launches Muse Wonder, its first frontier AI model family

Meta’s Superintelligence Lab has released Muse Wonder and Muse Wonder Full, a pair of larger AI models that TechCrunch describes as the company’s first attempt at a frontier model. The launch matters because Meta is moving beyond its lighter Muse Spark model and putting new options in front of developers, creators and businesses that already compare tools from OpenAI, Anthropic and Google.

The company says the Wonder models are designed for more demanding work, including complex tasks such as coding. Meta is also emphasizing memory: the ability for an assistant to retain and synthesize information a user provides over longer periods.

What Meta says is new in Muse Wonder

Muse Wonder and Muse Wonder Full sit above Muse Spark in Meta’s model lineup. Spark was positioned as a lightweight model built for low latency, while the new Wonder models are larger and, according to Meta, more capable.

In testing shared in the model card, Meta says Wonder and Wonder Full reached benchmark scores comparable to Fable 5.1 and GPT-6 Astra. The company’s launch post called the models a step on its “scaling ladder” and tied them to its broader push toward personal superintelligence.

For Emonarc readers, the key point is not just that Meta has a bigger model. It is that Meta now has a higher-end model tier to place against the assistants and APIs many teams already evaluate, including ChatGPT, Claude and Google Gemini. If your workflow depends on coding help, long-running research, or agents that need context from earlier interactions, Muse Wonder is a name to add to the comparison list.

For a broader buyer view of the existing assistant market, see our ChatGPT vs Claude vs Google Gemini guide.

Pricing and access are split between the two models

Meta is making Muse Wonder available through the Meta Model API. According to TechCrunch, access costs $4.00 per million input tokens and $20.00 per million output tokens. The report notes that this puts Wonder in roughly the same cost range as GPT-6 Sol or Opus 5.5.

Muse Wonder Full has a different path. Meta says Wonder Full will only be available to Muse Code subscribers and will not be served through the API. Both Muse Wonder and Muse Wonder Full will also power the Muse agent.

That split matters for small teams. If you build automations, customer support tools, internal assistants or content pipelines around APIs, only Wonder is directly relevant from the details Meta has shared. If you are using Muse Code, Wonder Full may matter more, but Meta has not described API access for that model.

Practical questions for teams to ask before testing include:

• Do you need API access, or are you working inside Muse Code?
• Are your tasks heavy enough to justify a larger model instead of a lower-latency one?
• Will the claimed memory improvements help your actual workflows, or are short, one-off prompts enough?

Meta is pushing memory as a differentiator

A major part of Meta’s pitch is what it calls “active memory.” The company is using the term for the models’ ability to remember and synthesize user-provided information over time.

Meta pointed to MemoryBench, a benchmark focused on continual learning and memory construction. According to the company’s testing, Muse Wonder scored higher on MemoryBench than Fable, Opus or Astra.

In a launch video, Meta research scientist Kavya Kopparapu said the team put significant work into making sure the model can remember what matters to users over “days, weeks, months and beyond.”

For creators and freelancers, this is the part to watch closely. Memory can be useful when an assistant needs to keep track of a brand voice, client preferences, project history or recurring tasks. But benchmarks do not always map cleanly to day-to-day work. Before moving sensitive or important workflows, teams should test whether the model remembers the right things, forgets what it should not retain, and handles changing instructions cleanly.

Meta’s emphasis also fits a broader industry move toward agents that can work across longer timelines. Google has also been pushing AI deeper into workplace tasks, including an enterprise Gemini agent for work across apps, while Anthropic has been expanding Claude’s agentic workflow capabilities, including dynamic workflows for parallel agents.

Why this launch matters for creators and small businesses

The Wonder release is Meta’s first frontier model launch since Alexandr Wang, founder of Scale AI, restructured the company’s AI efforts in August of last year, according to TechCrunch. Since that reorganization, Meta has released Muse Spark and the Muse agent, but had not made the kind of state-of-the-art model release associated with OpenAI and Anthropic.

Meta says the new models were trained at its Prometheus data center in central Ohio, which came online earlier this year.

For buyers, the immediate impact is more competition at the top end of the model market. That can be useful even if you do not adopt Meta’s models right away. New frontier releases tend to reset evaluation lists for coding, agent workflows, research, summarization and memory-heavy assistants.

The cautious approach is to benchmark Muse Wonder on your own tasks rather than relying only on published scores. Test it against your current model for the work that costs you time: code fixes, proposal drafts, client research, support responses, long project threads and internal documentation. Meta has made strong claims about memory and capability, but the value for a small business will come down to accuracy, cost, latency and how well the model fits existing tools.