Meta Releases Muse Wonder Models Built for AI Agents

Meta released two new AI models, Muse Wonder and Muse Wonder Full, on October 12, 2026. The company calls them its most capable models to date. It says they are built for agentic tasks, meaning multi-step work an AI carries out on a person's behalf. Meta's announcement says Muse Wonder is available now through the Meta Model API and Muse Code, its coding agent. Muse Wonder Full is available only to Muse Code subscribers. Meta says Muse Wonder will soon power experiences in Muse, the personal agent it introduced in September, starting with subscribers.

What did Meta release?

The models come from Meta Superintelligence Labs, the research group led by Alexandr Wang. TechCrunch called them Meta's first attempt at a frontier model, one that competes with the most capable systems. They are larger than Muse Spark, the lightweight model Meta introduced in April 2026. Meta says both were trained at its gigawatt-scale Prometheus data center. TechCrunch reported that the center is in central Ohio.

Artificial Analysis, an independent benchmarking firm, lists these details:

• Inputs and outputs: text, images, video and audio in; text out.
• Context window: 1 million tokens for each model. Tokens are the small chunks of text a model reads and writes.
• Muse Wonder price: $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens.
• Muse Wonder Full price: no per-token price. It is available only to Muse Code users.

Neither model can be downloaded and run elsewhere, VentureBeat reported. That is a shift from Meta's earlier Llama models, which organizations could run on their own hardware. The site's guide to open-weight and closed models explains the difference.

Users pay too. CNBC reported that Muse app users will need a monthly subscription of $20 or $100 to use Muse Wonder. Wonder Full requires a Muse Code plan of $50 or $200 a month, CNBC said.

Why does Meta pitch them at agents?

Meta's research post names two traits built into the models:

• Active Memory: carrying important context about a person across conversations, then pulling up the details a task needs.
• Active Focus: sticking with a hard goal by trying other approaches and working through blockers.

Meta says the model asks for clarification when needed. It also says the model checks in before taking important actions. Its demos show Muse Wonder making a running plan in a browser, building a 3D kitchen game and preparing a wedding seating chart.

Wang told CNBC the model was built so agents do not stumble on everyday tasks like booking dinner reservations. He said the team trained it to try other approaches instead of stalling. CNBC reported that its internal codename was Watermelon.

Meta also highlighted a memory test called MemoryBench. By Meta's own testing, Muse Wonder beat Anthropic's Fable and Opus models and OpenAI's Astra on it, TechCrunch reported.

How do the models score in outside tests?

Artificial Analysis combines ten tests into one score it calls the Intelligence Index. On October 12, it gave Muse Wonder Full 56 and Muse Wonder 52, both at maximum reasoning effort. Meta's previous model, Muse Spark 1.3, scored 48. The firm said the largest gains came on agent tests. On one of them, Terminal-Bench 4.0, Wonder Full reached 57%, up from 33% for Muse Spark 1.3.

VentureBeat listed rival scores on the same index. Among the models it listed, Anthropic's Claude Opus 5.5 scored highest, at 58. Anthropic's Claude Sonnet 5.5 tied Wonder Full at 56. OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Claude Fable 5.1 each scored 53.

The firm also rates coding agents, which pair a model with the tool that runs it. Muse Code with Wonder Full scored 68 on that index. Artificial Analysis listed Claude Code with Sonnet 5.5 at 68 and Opus 5.5 at 66.

Cost per task is another measure. Artificial Analysis estimated that one task on its index costs $4.52 with Muse Wonder. That compares with $5.98 for Claude Opus 5.5 and $3.26 for GPT-6 Astra. The firm said Muse Wonder's premium per-token prices and heavy input token use push its cost up.

Meta's model card shows wins and losses. VentureBeat reported that Wonder Full scored 57.5 on SWE-Marathon, a test of long-running coding work. Opus 5.5 scored 46.3 and GPT-6 Astra 42.5. On Terminal-Bench 4.0, both rivals scored higher than Wonder Full in Meta's table. VentureBeat noted that the model card mixes Meta's own tests, scores supplied by vendors and third-party results.

What are the safety and privacy questions?

Meta says a personal agent may work across a person's messages, calendars and documents. It says it trained the models not to share private information without permission and to turn down harmful requests. Its research post says Muse Wonder has Meta's strongest safeguards to date against chemical, biological, radiological, nuclear and cyber misuse. A full report with more safety tests will be available soon, Meta says.

Data use differs by product. VentureBeat reported that Muse Code subscriptions default to a Contributor mode. In that mode, Meta may use prompts, code and files to improve future models. Subscribers can turn it off, but only for future submissions. The pay-as-you-go API runs only in a Standard mode, the report said. Meta says content in that mode is not used for training.

The Muse agent itself has drawn complaints. CNBC reported that some users said on social media that it accessed personal messages without permission. Meta has disputed that. Wang said some users may have authorized communications earlier in a conversation and then forgotten. He told CNBC the team is watching for agent incidents and issues.

In September, Amazon blocked Muse from shopping on Amazon.com. Amazon told Adweek it was not told in advance that Muse would access its store. It also said the agent does not identify itself. Meta did not respond to Adweek's request for comment.

What comes next?

Meta introduced Muse on September 8, 2026, powered by Muse Spark. It runs on a dedicated cloud computer that Meta calls Muse Secure VM. A separate Sentinel agent must approve anything Muse sends to the internet. The AI agent launch tracker lists Muse among major agent launches.

Several points were still open as of October 16, 2026:

• The switch in Muse. Meta says Muse Wonder will soon power experiences in Muse, starting with subscribers. It has not given a date.
• The safety report. Meta says the full report will come soon, without a date.
• API access to Wonder Full. At launch, Meta had not announced a pay-as-you-go price for it, VentureBeat reported.
• An encrypted Muse. In September, Meta said it would introduce Muse Confidential VM later this year. It said that version would be encrypted with a key only the user holds.

For newcomers, the site's explainer on AI agents covers how agents like Muse plan and carry out tasks.