Meta has launched Muse Wonder and Muse Wonder Full, its most capable AI models to date, pushing the company deeper into the frontier-model market dominated by OpenAI, Anthropic and other major laboratories. The October 12 release combines long-running agent capabilities, persistent context and computer use with sharply different access models for the two variants. Meta shares rose about 4% in afternoon trading following the announcement.
According to Meta’s official Muse Wonder announcement, both models introduce Active Memory, which carries relevant information across conversations, and Active Focus, which allows agents to pursue goals, recover from errors and change approaches when tasks become difficult. The architecture is designed around agents that continue working across extended workflows rather than models limited to isolated prompts and responses.
Full Model Stays Behind Subscription Access
Distribution differs substantially between the two models. Muse Wonder is available through both the Meta Model API and Muse Code, with a one-million-token context window and API pricing of $4 per million input tokens, $0.20 for cached input and $20 per million output tokens. Muse Wonder Full is not available through Meta’s public pay-as-you-go API and instead requires one of the higher Muse Code subscriptions.
Full is currently included with the $50-per-month Power Usage plan and $200 Maximum Usage plan, while lower Muse Code tiers do not include Meta’s highest-capability model. That creates a practical split between the model developers can integrate directly through API calls and Meta’s strongest version. Access architecture, rather than model capability alone, will determine which businesses can deploy each system inside production workflows.
Meta’s benchmark results also require qualification. Muse Wonder, not Full, scored 85.6 on MemoryBench, narrowly above Full at 85.5. Wonder Full performed strongly on Harvey’s legal benchmark, Vals Finance Agent V2, DeepSWE and SWE-Marathon, but Claude Opus 5.5 remained ahead on several other evaluations. Independent evaluator Artificial Analysis scored Wonder Full at 56 and standard Wonder at 52, compared with 58 for Opus 5.5. Meta’s release establishes competitive frontier performance across several categories, not uncontested leadership across the benchmark suite.
Agentic AI Raises New Infrastructure Questions
The release has particular relevance for crypto companies experimenting with autonomous software. Blockchain developers are already building systems in which AI agents can transact, manage wallets or interact with financial applications, from multisig controls that give Bitcoin agents bounded authority to architectures where AI agents are expected to become major blockchain users. More capable long-running models increase what those agents can do, while making execution permissions, audit trails and data controls more consequential.
Meta’s own developer configuration illustrates one of those governance trade-offs. Standard mode does not use submitted prompts, code, files or model responses to train Meta models. Muse Code subscriptions, however, default to Contributor mode, which provides additional usage in exchange for allowing Meta to use that content for future model training; users can disable the setting. For organizations handling proprietary code or financial information, the distinction between model access and data-use configuration is operationally material.
The security dimension is developing in parallel. Crypto projects including Nethermind and ZEUS recently sought access to Anthropic’s frontier-model vulnerability scanner, demonstrating that increasingly capable AI systems are already entering blockchain security workflows. Meta similarly says Muse Wonder preserves advanced defensive cybersecurity capabilities while adding safeguards against malicious use. Frontier AI is therefore becoming both an automation layer and part of the security tooling surrounding digital-asset infrastructure.
Meta plans to bring Muse Wonder into its Muse personal agent next, initially for subscribers, while Muse itself is also scheduled to expand into the company’s AI glasses. The immediate launch, however, is primarily a developer and Muse Code release, and broader consumer deployment remains a separate rollout. For crypto and financial infrastructure teams, measurable adoption will depend on actual integrations, production workloads and whether the restricted Full model provides enough additional capability to justify its narrower distribution.
