Meta Unveils Muse Wonder AI Models, Claiming a Return to the Industry’s Frontier

Meta Platforms on Monday released Muse Wonder, a new family of artificial intelligence models that the company says returns it to the front rank of the AI industry after a 14-month overhaul of its research operation, setting up a direct challenge to OpenAI and Anthropic in the race to build software agents that act on users’ behalf.

The release includes two models: Muse Wonder, available to developers through the Meta Model API and the company’s Muse Code programming tool, and a larger Muse Wonder Full, offered to Muse Code subscribers on its $50- and $200-a-month plans. Meta said Muse Wonder will soon power its Muse personal-agent app, rolling out first to paying subscribers.

Chief Executive Mark Zuckerberg described the series as “a much larger and more advanced class of models that brings us to the frontier.”

Built for long-running agents

Meta is pitching the models at so-called agentic work, such as handling an insurance claim, reviewing a lease or building a working app. The company highlighted two capabilities it calls Active Memory, which carries context about a user across conversations, and Active Focus, which keeps the model working through obstacles and checking in before consequential actions.

The models were trained at Meta’s gigawatt-scale Prometheus data center, accept text, images, video and audio, and support a context window of 1 million tokens.

Meta AI chief Alexandr Wang told CNBC the goal was an agent that does not stall when a task hits a snag. “We specifically tried to train the model to be able to try other approaches instead of stalling and then checking back in with the user at any time,” he said.

Strong, but not on top

On Meta’s own scorecard, Muse Wonder Full led on long-horizon software engineering and legal-agent tasks, including the SWE-Marathon benchmark, where it scored 57.5 against 46.3 for Anthropic’s Claude Opus 5.5. Opus 5.5 led on several other measures, including OSWorld, Terminal-Bench and ProgramBench.

Independent testing placed the new models near, but not at, the top of the field. Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index, level with Claude Sonnet 5.5 and two points behind Opus 5.5, while the API-available Muse Wonder scored 52.

Pricing may temper the challenge. Muse Wonder costs $4 per million input tokens and $20 per million output tokens, identical to Opus 5.5, despite earlier suggestions from Wang that Meta would price “meaningfully below” its rivals.

Closed weights, safety caveats

The launch continues Meta’s retreat from the open-weight approach of its earlier Llama models; neither model’s weights are being released. In its model card, Meta said it could not rule out that Muse Wonder Full, without safeguards, would reach its high-risk threshold for chemical and biological misuse, and said it applies safety training and classifiers that bring residual risk to “moderate or lower.” A fuller safety report is promised soon.

The release lands amid heightened scrutiny of AI agents across the industry, as regulators and companies grapple with recent incidents in which autonomous systems took unintended actions on real-world websites and services.