Meta Launches Muse Wonder and Wonder Full Frontier AI Models

Meta has introduced Muse Wonder and Muse Wonder Full, its first frontier-class AI models developed under Meta Superintelligence Labs. The new systems target complex coding, long-running AI agents and personalised assistance, with Meta claiming performance comparable to leading models from OpenAI and Anthropic.

According to Meta’s official announcement, Meta offers Muse Wonder through the Meta Model API and Muse Code, while the more capable Wonder Full is limited to Muse Code subscribers. The models represent Meta’s biggest advancement since it reorganised its artificial intelligence operations under Alexander Wang.

Muse Wonder Targets Advanced Coding and AI Agents

Unlike the earlier Muse Spark model, which prioritised lightweight operation and low latency, Muse Wonder is designed for more demanding tasks. Meta says the new models deliver stronger performance in software development, professional knowledge work and multi-step agent workflows.

In its published model card, Meta compares the Wonder family with Claude Fable 5.1, Claude Opus 5.5 and GPT-6 Astra. The company reports competitive results across coding, reasoning and agent benchmarks, although some evaluations use internal testing methods or provider-reported scores rather than fully independent comparisons.

Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API, placing it in the premium pricing category for advanced AI models. Meta will not offer Wonder Full through the API, limiting access to subscribers of its coding service.

Active Memory Is a Major Focus

One of the model’s central capabilities is what Meta calls Active Memory, designed to retain important information across conversations and retrieve relevant details when needed. The company wants its AI assistants to maintain useful context over days, weeks, and months rather than requiring users to explain their preferences or previous instructions repeatedly.

Meta reports that Muse Wonder performs particularly well on its internal MemoryBench evaluation, which tests recall across long, complicated conversations. Separately, academic research on MemoryBench examines memory construction and continual learning in language-model systems, although that published benchmark is distinct from Meta’s internal test.

The models also introduce Active Focus, a capability intended to help AI agents persist through difficult tasks, explore alternative approaches and recover from obstacles. Combined with improved memory, this lets agents work on longer projects while retaining relevant instructions and context.

Meta’s Frontier AI Push Builds on Prometheus

The release follows Meta’s restructuring of its AI development efforts under Alexandr Wang, the former Scale AI founder who joined the company’s leadership in 2025. Meta previously released Muse Spark and its Muse agent, but Wonder marks a more substantial move toward the high-performance models developed by competing frontier AI laboratories.

Meta says it trained Muse Wonder and Wonder Full at its Prometheus data centre in central Ohio. The massive AI computing project, previously detailed in local reporting on the Prometheus facility, became operational earlier this year and supports the company’s expanded model-training efforts.

Muse Wonder is also expected to power Meta’s Muse AI agent, initially for subscribers, with broader applications planned for tasks requiring more advanced reasoning. The launch gives Meta a new flagship model family built around coding, sustained task execution, and long-term memory. However, its reported performance advantages largely rest on thecompany’ss published evaluations.