Meta Muse Wonder Arrives for Developers With 1 Million-Token Context and Agent-Focused AI Features
Meta has unveiled two new AI language models, Muse Wonder and the more powerful Muse Wonder Full, marking another major step in the company’s push toward agentic AI. Announced on Oct. 12, the models are built on the foundation of Muse Spark and are designed to help AI agents complete complex, multi-step tasks on behalf of users.
Instead of simply answering questions, these models are intended to take action. Meta says possible use cases include managing an insurance claim, reviewing and understanding a lease agreement, or creating a functional app from a user’s instructions. The goal is to make AI more useful in real-world workflows, especially where tasks require planning, persistence and access to relevant information over time.
One of the biggest highlights is the massive 1 million-token context window. This allows Muse Wonder and Muse Wonder Full to process and remember much larger amounts of information during a task, such as long documents, multi-part conversations, codebases or detailed project requirements.
Meta is also promoting two key agent-focused capabilities: Active Memory and Active Focus.
Active Memory is designed to help the AI carry important details across conversations, making it better at remembering context that matters. Active Focus, meanwhile, helps the model stay committed to a goal when a task becomes difficult. Instead of giving up quickly, the AI can try different methods, reassess its approach and ask for user confirmation before taking important steps.
Muse Wonder is available now for developers and builders through the Meta Model API and Muse Code, Meta’s coding agent. However, the more advanced Muse Wonder Full model is currently limited to Muse Code subscribers.
For everyday users, the rollout is still limited. Meta’s personal AI agent app, Muse, does not yet run on the new Muse Wonder model. The company says Muse Wonder will soon power experiences inside Muse and will first become available to subscribers, but it has not provided a specific launch date for free users.
Availability is also restricted by region. The Muse app launched in the United States in September and later expanded to Canada. Meta has not announced when the app will arrive in Europe, and the company has not provided a full country list for access to the Meta Model API.
Meta’s new models arrive as major tech companies continue to race toward more capable AI agents. These systems are increasingly being designed to move beyond chatbot-style conversations and into task execution, including writing code, organizing information, managing schedules and interacting with files or services.
Meta has shared benchmark results that present Muse Wonder Full as a highly competitive model. On the company’s own Personal Superintelligence Index, Muse Wonder Full reportedly scores 59.1, placing it ahead of Claude Opus 5.5 at 55.2 and GPT-6 Astra at 54.5.
However, Meta has not publicly explained exactly how this internal benchmark is calculated, which makes direct comparison difficult. The company has also noted that one of its coding comparisons, SWE-Marathon, was not a controlled test and that its model ran at maximum effort.
Independent evaluations paint a more cautious picture. On the Artificial Analysis Intelligence Index, which combines multiple evaluations into a weighted average, Claude Opus 5.5 currently leads with a score of 58. Muse Wonder Full scores 56, putting it close to the top and tied with Claude Sonnet 5.5.
The standard Muse Wonder model, which is the version available for developers through the API, scores 52. That places it just behind GPT-6 Astra and Claude Fable 5.1, suggesting that Meta’s latest model family is highly competitive, though not clearly dominant across outside tests.
Pricing is another important part of the launch. Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the API. That puts it in the same pricing range as Claude Opus 5.5 and below the listed cost of GPT-6 Astra, which is priced at $10 per million input tokens and $50 per million output tokens.
For developers building AI agents, these costs could matter significantly. Agentic systems often require many steps, multiple tool calls and long context processing, meaning token usage can grow quickly during complex tasks. A lower price may make Muse Wonder appealing for teams testing AI agents in coding, document analysis, workflow automation and business productivity tools.
For users who simply want to chat with an AI assistant, paid API access may not be necessary. Several free AI chatbot options continue to be available from competing services, including newer models in free consumer tiers.
Privacy and safety are also central concerns for agent-based AI. Because AI agents may interact with sensitive information such as email, calendars, contracts, documents and personal records, Meta says Muse Wonder and Muse Wonder Full were trained to avoid sharing private information without user permission.
The company says a full safety report will be released later, though it has not announced a publication date. Until then, developers and users will likely want clearer details about how the models handle personal data, permission boundaries, memory features and sensitive workplace information.
In Europe, users have the right to object to Meta using public Facebook and Instagram posts for AI training. This remains an important privacy option for people who do not want their public social media content included in AI development.
Overall, Muse Wonder signals Meta’s growing ambition in the AI agent space. With a 1 million-token context window, agent-focused memory and goal-tracking features, competitive benchmark results and relatively aggressive API pricing, Meta is positioning Muse Wonder as a serious option for developers building the next generation of AI tools.
The biggest questions now are when Muse Wonder will reach the consumer Muse app, how broadly it will be available worldwide and whether Meta’s promised safety report will provide enough transparency for users and businesses considering AI agents for sensitive tasks.
