Meta Launches Muse Wonder and Muse Wonder Full: Zuckerberg Says Meta Has Reached the AI Frontier

What happened

On October 12, Meta unveiled Muse Wonder and the more powerful Muse Wonder Full. In an official blog post, the company called them its most capable models yet, built for personal and professional agentic tasks. CEO Mark Zuckerberg wrote that Muse Spark was Meta Superintelligence Labs' first family of models, and that Wonder is a "much larger and more advanced class of models" that "brings us to the frontier." Meta shares rose more than 4% on the day, trading at $750.53 intraday.

The launch centers on two agentic capabilities: Active Memory, which carries what matters about a user into later conversations and brings back details relevant to the task at hand; and Active Focus, which keeps working toward the goal when a task gets hard — trying other approaches, recovering from errors, asking for clarification when needed and checking in before anything important. The models support roughly a 1-million-token context window, accept text, image, audio and video inputs, and output text. Demos showed the model building a personalized running plan across apps and a browser, and generating a playable 3D kitchen game from image inputs.

Muse Wonder is available now through the Meta Model API and Muse Code; the stronger Muse Wonder Full is limited to Muse Code subscribers, with no API pricing announced. The Muse personal agent app will get Wonder next, rolling out to paying subscribers first. Wonder is priced at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million — the same tier as Anthropic's Claude Opus 5.5 and Google's Gemini 4 Argon. The series was developed under the internal codename "Watermelon"; earlier in the day, Zuckerberg posted a photo of a watermelon, saying his own Muse assistant had sent it to him.

Key facts

1. Independent benchmarks: Artificial Analysis' Intelligence Index scored Wonder Full at 56, behind Claude Opus 5.5 at 58 and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53; the standard Wonder scored 52, tying GPT-6.1 Sol. Both are well above Muse Spark 1.3 (48), released six weeks earlier.
2. Meta's own benchmarks: On the long-horizon software benchmark SWE-Marathon, Wonder Full scored 57.5 against 46.3 for Claude Opus 5.5 in Meta's charts; it took top scores on professional legal and finance tasks, including Harvey's Legal Agent Benchmark (LAB). Meta also introduced three in-house evaluations: MemoryBench, FocusBench and AutomationBench-AA.
3. Training compute: Both models were trained at Meta's gigawatt-scale Prometheus data center.
4. Safety stance: Trained and released in line with Meta's Superintelligence Scaling Framework; Meta calls the family its "most aligned" models yet, designed to avoid sharing private information without permission and to refuse genuinely harmful requests. A full safety evaluation report is "coming soon."
5. Stock: Meta rose more than 4% on launch day (intraday $750.53); shares are up about 31% since the start of September, closing in on the August 2025 record high of $796.23.

Background

Superintelligence Labs was formed 14 months ago out of Zuckerberg's mid-2025 overhaul of Meta's AI efforts. In June 2025, Meta invested $14 billion in the San Francisco AI data company Scale AI and brought in its co-founder Alexandr Wang as chief AI officer to run the new division. In April, the lab released the first-generation Muse Spark, marking Meta's departure from the open-source Llama playbook.

Wonder's pricing puts Meta in the same tier as OpenAI and Anthropic: closed weights, with the flagship version subscription-only. Meta expects capital expenditures of $130 billion to $145 billion in 2026. In September, the hot launch of the Muse personal agent app drove Meta shares up 26.7% in a single month. Bank of America analyst Justin Post wrote in a research note last month that Superintelligence Labs' investments were translating into model performance and product innovation.

The launch was not without turbulence: The New York Times reported that Meta paused a pre-training stage in July, delaying the release. Since its debut, the Muse agent has also drawn multiple privacy-controversy reports, including changing a user's password without authorization and disclosing a user's home address; Meta responded that an agent "needs to have context about you and those you interact with" in order to be useful.