Meta Launches Muse Wonder, Its First Frontier-Class AI Models

Meta has released Muse Wonder and Muse Wonder Full, two new AI models built for what it calls "personal and professional agentic work." Independent testing puts the larger Full model second only to Anthropic's Claude Opus 5.5 on a leading intelligence index.

That's a big deal for Meta. After more than a year of rebuilding its AI operation, it finally has models that sit in the same conversation as the best from Anthropic, OpenAI, and Google.

How Meta got here

About 14 months ago, Mark Zuckerberg reorganized Meta's AI work and created Meta Superintelligence Labs (MSL), now led by Alexandr Wang. The first results were the Muse Spark models, which were solid but not frontier-leading.

Wonder is the next step up. Zuckerberg described it on X as "a much larger and more advanced class of models" than Muse Spark. Meta trained both models at Prometheus, its gigawatt-scale data center, using large-scale pretraining followed by reinforcement learning (training the model by rewarding good outcomes).

Why this matters: Meta has spent heavily on AI talent and compute. Wonder is the first real evidence that the spending is producing frontier-grade results.

What's new: memory and persistence

Meta is pitching two agent features above all.

Active Memory keeps relevant personal and professional context across conversations. If you told the assistant about a project last week, it can pull those details in when a new task needs them.

Active Focus helps the agent stick with hard goals. It tries other approaches when it gets stuck, works around blockers, asks for clarification, and checks with you before consequential actions.

Both models accept text, images, audio, and video, respond in text, and support a context window (how much the model can read at once) of about one million tokens.

How it scores

Artificial Analysis, an independent benchmarking firm, gave Muse Wonder Full a score of 56 on its Intelligence Index v4.3 and the standard Muse Wonder a 52. That's up from 48 for Muse Spark 1.3, released only six weeks earlier.

The biggest gains came on agent tasks. On Terminal-Bench 4.0, which tests command-line work, Wonder Full jumped from 33% to 57%. On Artificial Analysis's Coding Agents Index, Wonder Full scored 68, edging out Claude Opus 5.5 running in Claude Code at 66.

There are weak spots. Wonder Full's hallucination rate (how often it confidently makes things up) fell from 33% to 23%, but Google's Gemini 4 Argon is lower at 15%. Wonder Full also trails Opus 5.5 and GPT-6 Astra on factual knowledge tests.

The catch: you may not get the best one

Only the standard Muse Wonder is available through Meta's pay-as-you-go API, at $4 per million input tokens and $20 per million output tokens. That's roughly three times the price of Muse Spark 1.3.

Muse Wonder Full is limited to Muse Code subscribers, at $50 or $200 per month with usage caps. VentureBeat notes that Muse Code subscriptions default to a "Contributor" mode that lets Meta use your submissions to improve its models, though you can switch it off for future work.

Cost per task matters too. Artificial Analysis estimates an average task costs $4.52 on Muse Wonder, versus $3.26 on GPT-6 Astra and $1.99 on Gemini 4 Argon, largely because Wonder reads a lot of input tokens.

Who this affects

Developers get a new frontier option, especially for agentic coding. But the best model is locked behind a subscription, not an API.

Everyday Meta users will see Wonder arrive in the Muse assistant, with subscribers first.

Competitors now face a fourth serious frontier lab. Meta's distribution across Facebook, Instagram, and WhatsApp means even a near-best model can reach billions of people.

What insiders are saying

Artificial Analysis called the release "a material jump in intelligence over previous Meta models" and highlighted on X that Wonder and Wonder Full scored 52 and 56, "each a material gain over Muse Spark 1.3."

Privacy questions are already surfacing. Because these agents will touch your messages, calendars, and documents, Meta stresses discretion and misuse resistance, and says a fuller safety report is coming soon.

What's next

Watch for Wonder Full on the API. Developers will push hard for pay-as-you-go access to the stronger model.

Also watch the safety report. Meta calls Wonder its "most aligned" model, and independent testers will want to verify that.

The bottom line

Meta is back in the frontier race. Muse Wonder Full is close to the top, with standout agent and coding skills, but the best version is subscription-only and the API model isn't cheap. If you build agents, test Muse Wonder against your current model on real tasks, and check the data-sharing setting before you do.