Muse Wonder

Meta's newsroom formally announces Muse Wonder and Muse Wonder Full as its most capable models to date, optimized for personal and professional agentic tasks such as navigating insurance claims, reviewing leases, and building working applications. Both models build on groundwork laid with Muse Spark and target expert domains including law, finance, and health alongside frontier coding performance. Active Memory carries user-relevant details across conversations while Active Focus keeps agents working through errors, trying alternative approaches, and checking in before consequential actions. Safety training covers private information handling, refusal of harmful requests, and defenses against cyber attacks. Muse Wonder ships today on the Meta Model API and Muse Code, with Muse Wonder Full available to Muse Code subscribers.

WIRED reports that Meta's Muse Wonder and Muse Wonder Full signal the company's return to the AI frontier after a period of trailing rivals like Anthropic, OpenAI, and Google. The models are distributed through an API and Muse Code, with the best variant reserved to power features across Meta's platforms including the Muse personal agent. Meta declined to disclose training or architecture details. According to WIRED, the largest Muse Wonder variant outperforms Anthropic's Claude Fable 5.1 and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. Context notes Meta's $14 billion Scale AI investment and CEO Alexandr Wang's leadership of Meta Superintelligence Labs, framing the release as payoff for Zuckerberg's year-long AI overhaul.

Meta's AI Research blog today introduces the Muse Wonder series, headlined by Muse Wonder and Muse Wonder Full as Meta's most capable models targeting personal and professional agentic work. The series combines Active Memory for cross-conversation context and Active Focus for sustained task execution, trained at Meta's gigawatt-scale Prometheus data center. The post positions these models as a step toward personal superintelligence. Muse Wonder is live now through the Meta Model API and Muse Code, with Muse Wonder Full available to Muse Code subscribers. Both variants will eventually power experiences in the Muse personal agent, starting with subscribers, and support artifact generation requiring higher intelligence. The blog reports state-of-the-art results across personal agents, professional agents, memory, and coding benchmarks, though the benchmark section is truncated in the excerpt.

Handy AI's newsletter catalogs Muse Wonder (key muse-wonder-1.0) at $4 input and $20 output per million tokens, with cached input at $0.20 and a 1,048,576-token context window. The model accepts text, images, video, audio, and PDF inputs, returns text, and supports reasoning effort from minimal to max. Pricing and cache rates match Claude Opus 5.5 exactly. Muse Wonder Full is gated to Muse Code on $50 Power and $200 Maximum Usage subscription tiers and is not reachable via API even at maximum reasoning. The newsletter positions the base model behind Claude Opus 5.5 and Sonnet 5.5, level with GPT-6 Astra but ahead on long-running coding. Recommended use cases include multi-hour coding agents and legal or document-heavy agent work.