Meta Launches Muse Wonder AI Models to Rival OpenAI and Anthropic

Meta Launches Muse Wonder AI Models to Rival OpenAI and Anthropic

Meta has spent months rebuilding its AI strategy. Now, it is making a much bigger statement.

On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, two new AI models developed by its Superintelligence Labs. The company describes them as its most capable models yet, designed to handle demanding professional tasks, complex coding, and AI agents that can work towards goals over time.

The launch matters because Meta is no longer focusing only on lightweight AI models that respond quickly. It wants to compete at the frontier of AI development, where companies like OpenAI and Anthropic are pushing models to solve harder problems and complete more complicated tasks.

But what makes Muse Wonder different from another chatbot? The answer comes down to three things: memory, persistence, and the ability to take action.

Meet Muse Wonder and Muse Wonder Full

Meta introduced two versions; each aimed at a different level of work.

Muse Wonder is available through Meta's Model API and Muse Code, giving developers access to its capabilities for building software and AI-powered applications. Muse Wonder Full is the more advanced version, designed for demanding workflows. However, it is currently available only to Muse Code subscribers, rather than through the public API.

According to Meta, both models are built to handle tasks that require more than a quick response. That could mean working through a large software project, reviewing complicated documents, or navigating a multi-step professional task.

For example, instead of simply suggesting how to build an app, an AI agent powered by these models could work through the code, make changes, run tests, and recover when something goes wrong.

That is the direction Meta wants to take AI: from giving instructions to helping carry them out.

The big feature: AI that remembers you

One of Meta's biggest selling points is a capability it calls Active Memory.

Think about how frustrating it can be to explain your preferences, goals, or project details to an AI assistant repeatedly. You might discuss a business idea today, revisit it next week, and still need to explain the background again.

Meta wants to reduce that friction.

Active Memory is designed to help Muse Wonder carry important information across conversations and retrieve it when it becomes relevant to a task. The goal is for an assistant to remember what matters to you over time, rather than treating every conversation as a fresh start.

Meta also highlights Active Focus, which helps the models keep working when a task becomes difficult. They can try alternative approaches, adapt to new information, recover from errors, and ask for clarification when necessary.

In practical terms, the two capabilities are intended to work together: memory helps the model understand the context, while focus helps it continue towards the goal. This could be useful for developers managing long-running projects, professionals working with large amounts of information, and people who want an assistant that understands their ongoing priorities.

Of course, how consistently these features work in everyday use will matter more than the names Meta gives them.

Can Muse Wonder compete with OpenAI and Anthropic?

Meta says its new models achieve state-of-the-art results across several areas, including coding, memory, and personal and professional AI agents. Its published comparisons put the Wonder models alongside systems such as OpenAI's GPT-6 Astra and Anthropic's Fable 5.1 and Opus 5.5 on selected evaluations.

Meta also highlights MemoryBench, a benchmark designed to assess AI memory and continual learning. The company says Muse Wonder performs strongly on this test.

These are notable claims, but there is an important distinction: benchmark results do not automatically mean one model is better at everything. Results can vary depending on the task, testing method, and amount of reasoning or computing time allowed. Independent testing and real-world use will help establish how the models compare across different workloads.

Still, the release signals that Meta wants to compete directly with leading AI developers rather than remain focused mainly on smaller, faster models.

How much does Muse Wonder cost?

Meta has announced the following pricing for Muse Wonder through its Model API:

Input: $4 per million tokens.

Output: $20 per million tokens.

Muse Wonder Full: Available through eligible Muse Code subscriptions, not the public API.

Tokens are the units AI models process when reading prompts and generating responses. Your actual cost will depend on how much information a task uses and how much the model produces.

For developers, that means the price of a model is only part of the decision. A model that completes a task with fewer retries or less manual intervention could offer better value, even if its per-token price is higher than a smaller alternative.

Muse Wonder is also set to power more features inside Meta's Muse personal AI agent, with the rollout beginning with subscribers.

Why this launch matters for Meta

This release follows a major overhaul of Meta's AI efforts, which began around 14 months ago. The company previously introduced Muse Spark, a lighter model intended for responsive, everyday use, and later launched Muse, its personal AI agent.

Muse Wonder represents a step towards a more capable system that can handle complex tasks over longer periods.

The models were trained at Meta's Prometheus data centre in central Ohio, part of the infrastructure the company is investing in to support its AI ambitions.

Meta is also emphasizing safety. It says the Wonder models were trained to protect private information, resist harmful requests, and support defensive cybersecurity work while applying safeguards against misuse. The company says additional safety evaluation details will be released separately, so those protections still need to be assessed alongside independent evidence.