Meta launches Muse Wonder, its first frontier AI model, betting on memory to beat OpenAI

Meta has finally shipped the model it has been promising since it rebuilt its AI group last year. Meta Muse Wonder, released on Monday, Oct. 12, along with a bigger sibling called Muse Wonder Full, is the company's first true frontier model, and Meta says it can stand next to the best systems from OpenAI, Anthropic and Google. The pitch is not only raw intelligence. Meta is betting that memory, the ability of an AI to remember what matters to you across conversations, is what will make personal agents useful.

The launch matters because Meta spent most of the past year as the big AI lab without a top tier model. It released the lighter Muse Spark and a Muse agent, but nothing that competed at the high end. With Wonder, it is back in the race, and it is pricing the model to compete directly with the most capable paid options on the market.

What Meta announced

According to Axios, the new models were known internally by the code name Watermelon. There are two versions:

• Muse Wonder, available through the Meta Model API and through Muse Code, Meta's coding product.
• Muse Wonder Full, a larger version reserved for Muse Code subscribers. TechCrunch reports it is not offered through the API.

Both models will power the Muse agent, Meta's assistant product. Axios says Wonder will drive future experiences in the Muse assistant, starting with paid subscribers.

Meta trained the models at Prometheus, its gigawatt scale data center in central Ohio, which came online earlier this year. That detail matters more than it looks. Meta has spent heavily on compute, and Wonder is the first big public result of that spending.

Pricing

TechCrunch reports that Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API. The outlet notes that this is roughly in line with GPT-6 Sol from OpenAI and Claude Opus 5.5 from Anthropic. In other words, Meta is not trying to undercut rivals on price. It is asking developers to pay premium rates and judge the model on quality.

That stands in contrast to the recent rush toward cheaper models, such as Anthropic's Claude Haiku 5.5 price cut and Microsoft's low cost Decision-1 model.

How Meta Muse Wonder stacks up

Meta published benchmarks showing Wonder topping or matching the latest models from OpenAI and Anthropic on personal and professional agent tasks, memory and coding, Axios reports. TechCrunch, citing Meta's model card, says Wonder and Wonder Full score in the same range as "Fable 5.1" and "GPT-6 Astra."

On one test, Meta claims a clear lead. TechCrunch says that on MemoryBench, a benchmark for continual learning and memory, Muse Wonder beats Fable, Opus and Astra. Neither outlet published specific scores, so the size of the gap is not clear yet.

Meta also says the models do especially well on unfamiliar tasks and have specialized strength in law, finance and health, according to Axios.

A note of caution on benchmarks

Company benchmarks are a starting point, not a verdict. Every major lab publishes numbers that put its new model in a good light, and independent testing often tells a more mixed story. Axios flagged the same open question: whether the performance claims hold up when outside researchers run their own tests. Developers who are thinking about switching should run Wonder on their own workloads before committing.

Active Memory and Active Focus

The two features Meta is leaning on hardest are called Active Memory and Active Focus.

Active Memory keeps important information about a user across conversations and pulls it back up when it becomes relevant, Axios reports. In a launch video cited by TechCrunch, Meta research scientist Kavya Kopparapu said the team worked to make sure "your model can remember what's important to you."

Active Focus is aimed at agents. It keeps an agent working toward a goal, trying different approaches and adjusting when it runs into obstacles. That is the kind of persistence that separates a chatbot that answers a question from an agent that can finish a multi step job. If you want a primer on what agents can and cannot do today, our guide to AI agents explained covers the basics.

Meta Superintelligence Labs head Alexandr Wang told Axios that "memory is incredibly important for personal agents." He also called Wonder and Wonder Full "true frontier models, competitive with the best models in the world."

Meta's own blog post describes the release as "a major leap toward personal superintelligence" and "the next step on our scaling ladder," according to TechCrunch. Personal superintelligence is the long term goal Mark Zuckerberg has set out: AI that understands each person's goals and preferences.

Why this matters

For the past year, the frontier race has mostly been a three way contest between OpenAI, Anthropic and Google. Meta has the money, the data centers and billions of users, but it lacked a model at the top of the leaderboard. Wonder is its attempt to change that.

There are a few reasons the launch is worth watching:

• Competition on agents. OpenAI recently pushed GPT-6 to everyone in ChatGPT, as we covered in our report on the GPT-6 Intelligent UI rollout. Meta is now pitching memory as its edge in the same agent market.
• Distribution. Meta can put a model in front of more people than almost anyone. How quickly Wonder reaches Facebook, Instagram and WhatsApp users is still an open question, Axios notes.
• Open source. Meta built its AI reputation on open Llama models. It has not said whether it will release weights for any Wonder model. Wang told Axios that Meta recently released the weights for Muse Glimmer and that "we're still committed to open source."

Who is affected

Developers get another premium option through the Meta Model API. At $4 and $20 per million tokens, the decision will come down to whether Wonder does better on their specific tasks than the models they already use.

Muse Code subscribers get access to Wonder Full, the most capable version, which is not available anywhere else.

Everyday users will mostly meet Wonder through the Muse assistant. Paid subscribers come first, so free users may wait.

Businesses in regulated fields such as law, finance and health should note Meta's claims of specialized strength there, but should also test carefully before relying on any model for high stakes work.

Safety and privacy questions

Meta says it strengthened safeguards against cyber and biological misuse and added protections for personal information, Axios reports. A model card with some details was due Monday, with a fuller safety report to follow.

Memory raises its own concerns. A model that remembers personal details across conversations is more useful, but it also holds more sensitive data. Axios pointed out that agents built on personal memory raise privacy and security questions. Users should look for clear controls that show what the assistant has stored and let them delete it.

What to watch next

• Independent benchmarks. Expect outside testers to publish scores in the coming days. Those will show whether Wonder truly matches the best from OpenAI and Anthropic.
• The full safety report. Meta promised more detail beyond the model card.
• Open weights. Any decision to release Wonder weights would be a major shift in the market.
• Consumer rollout. Watch for Wonder to appear in Meta's main apps and for any change to the paid tier requirement.

For more coverage of the model race, visit our AI news section or read about OpenAI shelving GPT-6.1 Astra and shipping Sol.

Frequently asked questions

What is Meta Muse Wonder?

Muse Wonder is Meta's first frontier AI model, released Oct. 12, 2026. It comes in two versions, Muse Wonder and the larger Muse Wonder Full, and focuses on memory, agents and coding.

How much does Muse Wonder cost?

Through the Meta Model API, Muse Wonder costs $4 per million input tokens and $20 per million output tokens, according to TechCrunch. Wonder Full is limited to Muse Code subscribers.

Is Muse Wonder open source?

Not so far. Meta has not said whether it will release Wonder weights, though Alexandr Wang said the company is still committed to open source.

Is Muse Wonder better than GPT-6 or Claude?

Meta says it matches or beats rival models on several benchmarks, including a lead on MemoryBench. Independent testing has not yet confirmed those claims.