Meta introduces Muse Wonder AI models as Zuckerberg rejects calls to slow development

Meta’s announcement, dated October 12, 2026, presents Muse Wonder and Muse Wonder Full as the company’s most capable models to date.[2] Meta says the models build on the groundwork laid by Muse Spark and deliver state-of-the-art performance on personal and professional agentic tasks.[2] The launch follows an overhaul of Meta’s AI efforts that the company says began fourteen months earlier; Meta describes its work as having moved from rebuilding foundations to pushing the frontier.[2] Separately, Zuckerberg said Meta had caught up with rivals at the AI frontier and rejected calls to slow AI development.[1]

Developers and builders can access Muse Wonder through the Meta Model API and Muse Code, according to Meta.[2] Muse Wonder Full is available to Muse Code subscribers.[2] Meta says Muse Wonder will soon power experiences in Muse, with the rollout starting with subscribers.[2] Meta also says the models offer capabilities in expert domains including law, finance and health, alongside frontier coding performance.[2]

Meta says Active Memory carries relevant context across conversations and retrieves details for the task at hand.[2] The company says Active Focus helps models persist through difficult tasks, try other approaches, recover from errors and ask for clarification when needed.[2] Meta says it trained the models not to share private information without permission, assist with harmful acts or undermine cybersecurity defenses.[2] A full report with additional safety evaluations was not yet available in Meta’s statement; the company said it would be available soon.[2] These are company descriptions, not a substitute for examining the forthcoming evaluations against a team’s own requirements.

Meta says Muse Wonder Full outperforms an Anthropic model and ranks just below another on the Artificial Analysis Intelligence Index.[4] It also says Muse Wonder outperforms models from OpenAI and Anthropic on some specialized benchmarks.[4] Those statements describe selected comparisons and do not establish performance across every task a buyer may care about. CNBC reported that some Muse users alleged on social media that an agent accessed personal messages without authorization; Meta disputed the claim.[5] For teams considering deployment, the benchmark claims, forthcoming safety report and disputed user allegation are distinct evidence to weigh rather than a single verdict on readiness.