Meta Releases Muse Wonder and Says Its Top Model Ranks Just Below Anthropic's Best

Meta released Muse Wonder and Muse Wonder Full on Monday, the first models it has said are meant to compete with frontier models from Anthropic and OpenAI. Meta says the larger model ranks just below Claude Opus 5.5 and above Claude Fable 5.1 on the Artificial Analysis Intelligence Index. The comparison table is Meta’s own compilation.

Benchmark comparisons

The index is a composite of evaluations that Artificial Analysis runs independently.

On SWE-Marathon, which tests extended software projects, Abundant AI’s results in Meta’s launch materials put Muse Wonder Full at 57.5%, against 46.3% for Claude Opus 5.5. The model card says the row “is not a controlled common-harness comparison.” Meta also ranks Wonder Full first on its own Personal Superintelligence Index, an internal composite of seven evaluations.

Meta’s model card says it used third-party or provider-reported results where available, including Artificial Analysis for its index and Abundant AI for SWE-Marathon, and its own runs elsewhere. The scores appeared in Meta’s launch materials. It calls internally generated competitor runs “best-effort” and says prompts, tools and runtime “may not be optimally tuned for proprietary models.”

Safety assessment

Meta’s model card says that without mitigations, it does “not rule out” Muse Wonder Full reaching its high-risk threshold in the chemical and biological domain. A model reaches that threshold “when its deployment could substantially contribute to a threat scenario we have identified as severe.” The company applied safety post-training and system-level classifiers. With those safeguards, it rates residual risk “moderate or lower” across domains.

Meta says a full report with additional safety evaluations will be available “soon,” but gave no date.

Meta has not said whether it will open-source the Wonder models. “We're still committed to open source,” chief AI officer Alexandr Wang said.

Memory and sustained work

Meta says Active Memory retains important information about users across conversations and retrieves details relevant to a new task. Meta says Active Focus keeps an agent working toward a goal by trying alternative approaches and adapting when obstacles arise.

“When you do real-world tasks, you’re going to get roadblocks and you’re going to get things that come up or various setbacks,” Wang said.

The models, codenamed Watermelon, were trained at Meta’s Prometheus data center. Wang, formerly of Scale AI, became chief AI officer in June 2025.

Muse agent record

Some Muse users have said the agent accessed personal messages without authorization. Meta disputes those complaints. Wang said some users had authorized communications earlier in a conversation “and then may have forgotten about that.”

In a hands-on account published the morning of the release, a reporter using the free tier got a restaurant reservation above the requested budget and had to submit Muse’s airline refund drafts manually. United Airlines, which prohibits bots, would not accept the reporter’s manually entered login in the browser window inside the Muse chat, and the reporter handled the task without Muse. The account did not name the model and was not a test of Muse Wonder.

Muse launched September 8 and reached 5 million downloads in 22 days, Sensor Tower said.

Access and prices

Muse Wonder is available now through the Meta Model API and Muse Code. Muse Wonder Full requires a Muse Code subscription at $50 or $200 a month.

OpenRouter’s listing prices Muse Wonder at $4 per million input tokens and $20 per million output tokens, with a context window of 1 million tokens. Wang said API pricing would be “meaningfully below” leading OpenAI and Anthropic models.

Meta says Muse Wonder will reach the Muse app “soon,” subscribers first, but has given no date. Its paid plans currently cost $20 or $100 a month.

Wang said, “I would tell folks they should try it out and judge for themselves.”

Google Launches Gemini Agent That Gives AI Coworkers Their Own Workspace Email

Google Cloud launched the Gemini agent on Thursday, October 8, letting companies give an AI coworker its own company email address, calendar and directory entry. The agent takes objectives, plans work

The Implicator

Manus Gives Cue Agents Their Own Phone Numbers and Wallets After Meta Split

Manus launched Manus 2.0 and Cue on Monday, Sept. 28, giving each Cue agent its own email address, phone number, wallet and computer. The standalone app lets an agent send messages, make payments insi

The Implicator

Nvidia Says Its New Agent Safety Platform Could Have Stopped the Hugging Face Breach

Nvidia launched its Open Agent Safety Platform on Monday and said it could have stopped OpenAI’s agents from breaching Hugging Face in July. The system pairs an open-source runtime that limits agent a

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai