Meta releases Muse Wonder frontier model family with long-running agent focus

Meta has rolled out its first family of frontier AI models, named Muse Wonder and Muse Wonder Full. The release marks a change in strategy for the company, moving away from its previous open-source approach to offer proprietary, closed-weight systems built for complicated, multi-step tasks.

The smaller Muse Wonder model is available through the Meta Model API at US$4 per million input tokens and US$20 per million output tokens, alongside a context window of one million tokens. The more capable flagship, Muse Wonder Full, is not available on the public programming interface and is restricted to subscribers of the Muse Code platform, with monthly plans priced at US$50 and US$200. Both systems are also set to back Meta's consumer and professional Muse assistant tools.

Retaining context and handling long tasks

Meta is focusing on two technical capabilities in this release: persistent memory and task recovery. The memory feature is built to store and connect user information and instructions across days, weeks, and months, rather than resetting after an individual session ends.

The second feature addresses common failure points in automated agent workflows. When an automated process hits an error, like a failed tool call or an unexpected website response, the model is designed to try alternative methods rather than stopping or looping indefinitely. Early benchmark figures show the systems perform strongly on extended coding, legal document review, and multi-step administration, matching competing offerings from OpenAI and Anthropic on several reasoning measures.

What this means for Australian businesses

• Longer memory reduces repetitive setup. For businesses using AI for customer correspondence or operational research, models that remember background details across weeks can reduce the need to repeatedly paste the same company context into prompts.
• Autonomous workflows get more dependable. Small companies trying out autonomous agents for data entry, invoicing follow-ups, or code maintenance will see fewer breaks caused by minor software hiccups.
• Vendor lock-in risk increases. Because Meta chose closed weights for this lineup rather than open releases, Australian teams cannot host the system on their own local infrastructure. Companies must weigh convenience against relying on a single cloud vendor for proprietary data.
• Pricing is predictable. The API costs match the current standard rates of rival enterprise systems, making budget planning straightforward for teams already paying for automated services.