14/10/2026
Meta is making another push toward the front of the AI race.
The company introduced Muse Wonder and Muse Wonder Full on October 12, describing them as its most capable artificial intelligence models to date.
Both are designed for agentic tasks that require AI to plan, work across multiple steps and recover when something goes wrong. Meta says potential uses range from interpreting leases and navigating insurance claims to analyzing finances and building complete applications.
The models introduce what Meta calls Active Memory, which carries relevant information between conversations, and Active Focus, which helps the AI continue working toward a goal, try different approaches and request clarification when necessary.
Muse Wonder is available through the Meta Model API and the company’s Muse Code development tool. It supports a context window of more than one million tokens and costs $4 per million input tokens and $20 per million output tokens through the API.
The more capable Muse Wonder Full is currently limited to subscribers using Muse Code. Meta also plans to use Wonder to power its consumer-facing Muse AI agent, beginning with subscribers.
Meta says its internal and third-party benchmark results place the models alongside leading systems from OpenAI and Anthropic. However, broader independent testing will be needed to establish how consistently they perform in real-world use.
Could Meta’s newest models close the AI gap with OpenAI and Anthropic?
