Meta introduced Muse Wonder and Muse Wonder Full on Monday, positioning the two models as its most capable systems yet for personal and professional AI agents. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Meta plans to use Wonder in its Muse assistant, starting with subscribers.
Memory and focus for longer tasks
Meta says the models combine two features aimed at keeping agents useful over time. Active Memory is designed to carry relevant information across conversations and retrieve it for a new task. Active Focus is meant to keep an agent working when it hits an obstacle, including trying another approach rather than immediately handing the problem back to the user.
The company claims the Wonder models reach state-of-the-art results on benchmarks covering agent work, long-term memory and coding. Those scores put Meta back in the conversation with leading systems from Anthropic and OpenAI, according to WIRED's assessment. Benchmark results, however, do not guarantee the same performance on messy real-world work.
Paid access and a proprietary release
The standard Muse Wonder model costs $4 per million input tokens and $20 per million output tokens through Meta's API. The more capable Full version is not offered through that API and instead sits behind paid Muse Code plans.
That distribution marks a contrast with Meta's earlier Llama strategy. The company helped popularize downloadable, open-weight models, but neither Wonder model is being released with public weights, VentureBeat reported. Meta says both were trained at its gigawatt-scale Prometheus data center. It has not disclosed detailed training methods, leaving independent testing to show how much the benchmark gains carry over to everyday agent tasks.
