Meta's New Muse Wonder Models Challenge OpenAI, but Access Comes With Limits

Meta is making another push into the frontier AI race, introducing two models designed to compete with OpenAI and Anthropic in coding, complex reasoning, and autonomous tasks.

The company announced Monday that Muse Wonder and Muse Wonder Full are its most capable AI models to date. Built for personal and professional AI agents, the models are designed to handle tasks ranging from software development to document analysis and financial research.

The launch follows about 14 months of work to rebuild Meta's AI capabilities under Meta Superintelligence Labs, led by Chief AI Officer Alexandr Wang. The company developed the models using its Prometheus data center and techniques spanning pretraining and reinforcement learning.

The release also marks a change in Meta's approach. Unlike its earlier open-weight Llama models, neither Muse Wonder model is available for public download and self-hosting.

Memory and persistence drive the new models

Meta's new models feature two capabilities designed to make AI agents more useful for extended tasks.

Active Memory lets the models retain relevant information across conversations and retrieve it when needed. Active Focus helps them work through obstacles, try alternative approaches, recover from errors and request clarification when necessary. The models are also designed to seek confirmation before important actions.

These capabilities target a common weakness in AI agents: completing a multistep task requires more than producing a convincing answer. An agent must track requirements, use tools correctly and respond when its original plan fails. Meta says the models also deliver advanced coding performance and specialized knowledge in areas such as law, finance and health.

Benchmark results put Meta back in contention

Independent evaluations from Artificial Analysis give Muse Wonder Full a score of 56 on its Intelligence Index, tying it with Anthropic's Claude Sonnet 5.5. Muse Wonder scores 52, while OpenAI's GPT-6 Astra scores 53.

The results suggest Meta has become more competitive with established frontier AI providers. However, the models do not lead every evaluation, and benchmark performance does not guarantee reliable results in real business environments.

Benchmark rankings also tell only part of the story. For businesses evaluating AI agents, reliability, cost, data handling, and performance on real-world tasks may matter just as much as a model's overall score.

Access comes with pricing and privacy limits

Muse Wonder is available to developers through Meta's Model API and Muse Code programming assistant. The API costs $4 per million input tokens and $20 per million output tokens, with discounted rates for cached inputs.

Muse Wonder Full is restricted to Muse Code subscribers, with subscription plans priced at $50 and $200 per month. Meta has not announced API access for the more capable model. Both models will eventually power features in Meta's Muse personal AI app, initially for subscribers.

What the launch means for businesses and users

For businesses, Muse Wonder adds another option for automating software development, document analysis, and multi-step workflows. Its pricing gives developers a concrete basis for testing it against competing models, although total costs will depend on how much reasoning, tool use and repeated processing each task requires.

Privacy will be equally important. Meta says the models are trained to protect private information and resist harmful requests. However, VentureBeat reports that Muse Code subscriptions default to a Contributor mode that allows submitted material to be used to improve future models. Subscribers can disable that setting, while the standard API does not use submitted content for training.

The different access options also come with different data-handling considerations. According to VentureBeat, Muse Code subscriptions default to a Contributor mode that allows submitted material to be used to improve future models, although subscribers can disable the setting. The standard Model API, by contrast, reportedly does not use submitted content for training.

Consumers may benefit when Muse Wonder reaches the Muse app, particularly for tasks that require remembering previous conversations or recovering from setbacks. But persistent memory and access to personal messages, calendars and documents also raise questions about permissions and data handling.

Meta says a fuller safety evaluation report is forthcoming. Until more evidence is available, users and organizations should verify consequential outputs and carefully review access permissions before allowing an AI agent to act on their behalf.

Editor’s note: This article originally appeared on our sister publication, TechRepublic.