Meta Launches Muse Wonder: Claims Frontier Models for Personal and Pro Agents
Meta Superintelligence Labs ships Muse Wonder and Muse Wonder Full, saying they match or beat OpenAI and Anthropic on agent, memory, and coding work—and will soon power the Muse assistant for paid users first.
By 1EyeAI Desk ·
Update (Oct. 13, 2026, 8:05 PM ET): When we published, no outside group had tested Muse Wonder. Artificial Analysis and Vals AI have now released independent benchmark results, which we’ve added below. Meta has also published the model card this story said was coming. See our corrections log.
Meta on Monday (Oct. 12, 2026) introduced the Muse Wonder series—Muse Wonder and the stronger Muse Wonder Full—as its first models it publicly frames as frontier-class after a year of rebuilding under Meta Superintelligence Labs. In Meta’s own research post, the company says the pair deliver state-of-the-art results on personal and professional agent tasks, memory, and coding, and were trained at its gigawatt-scale Prometheus data center.
Access is live now: Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is for Muse Code subscribers. Meta says Wonder will soon power experiences inside the Muse personal agent, starting with subscribers, and later feed specialized models aimed at broader Muse users. Meta AI chief Alexandr Wang told Axios the models are “true frontier models, competitive with the best models in the world,” especially on agentic personal and professional work. The Verge reports Meta positions them against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 on internal benchmarks—those comparisons are Meta’s own evaluations; see the independent results below.
Two product ideas sit at the center. Active Memory is meant to carry what matters about a user across conversations and pull it back when a new task needs it. Active Focus is meant to keep an agent grinding toward a goal—trying other approaches, adapting when blocked, and checking in before important actions. Meta also highlights domain work in law, finance, and health, plus long-horizon coding tests such as SWE-Marathon. Meta has since published its model card (dated Oct. 12) and has said a fuller safety report will follow; Meta says Wonder has its strongest safeguards yet against CBRN and cyber misuse while still allowing defensive cyber use under its scaling framework.
Codename context (labeled): executives had been teasing “Watermelon” on social media ahead of the drop; Axios confirms that was the prior internal name for this family.
Independent results (added Oct. 13): Artificial Analysis scores Muse Wonder Full 56 and Muse Wonder 52 on its Intelligence Index v4.3 (at max reasoning effort), up from 48 for Meta’s previous Muse Spark 1.3. On Vals AI’s Vals Index (updated Oct. 13), Wonder Full ranks #5 at 64.57%, behind Gemini 4 Argon and three Claude models but ahead of GPT-6 Astra (63.13%); Muse Wonder ranks #10 at 61.06%.
What this means for you: If you build agents or ship coding tools, put Muse Wonder on a short bake-off against Claude and GPT on your multi-step workflows—especially anything that needs durable memory—before you rewrite stack defaults. Independent tests put it in the top tier but not clearly ahead of the best Claude models, so judge it on your own work. Privacy-wise, memory-heavy agents that touch calendars, mail, and docs raise real data-risk questions; keep permissions tight. Not financial advice—model launches move stock narratives, not your product roadmap by themselves.