Meta Unveils Muse Wonder, Its Most Powerful AI Model to Date

Meta has launched two new artificial intelligence models called Muse Wonder and Muse Wonder Full. The company announced them on October 12, 2026. Meta says these are its most capable models to date. They were built by Meta Superintelligence Labs, led by Alexandr Wang.

What the new models can do

Muse Wonder Full is the stronger of the two. It scored 56 points on the Artificial Analysis Intelligence Index, a well-known AI ranking. That puts it second overall, behind only Anthropic’s Claude Opus 5.5 with 58 points. The smaller Muse Wonder scored 52 points, just behind OpenAI’s GPT-6 Astra and Google’s Gemini 4 Argon. Both models accept text, images, video, and audio. They can process up to one million tokens at once.

Meta gave the models two special skills. The first is called Active Memory. It helps the model remember important facts about the user over days and weeks. The second is called Active Focus. It helps the model stick with hard problems, try different approaches, and ask follow-up questions when it gets stuck. These skills make the models good at agent work: long tasks that use tools across many apps.

Strong results in coding tests

The new models did very well in coding benchmarks. Muse Wonder Full scored 57.5 percent on SWE-Marathon, a test that checks how AI handles long software engineering projects. Claude Opus 5.5 scored 46.3 percent on the same test. GPT-6 Astra scored 42.5 percent. On another coding test called DeepSWE v1.1, Muse Wonder Full reached 79 percent, ahead of both rivals. Meta says the models are also strong in law, finance, and health tasks.

Price and availability

There is one big change from Meta’s past. The Wonder models are not open. Unlike Meta’s older open Llama models, the Wonder models are closed like those of OpenAI and Anthropic. Muse Wonder is available through Meta’s Model API. It costs 4 dollars per million input tokens and 20 dollars per million output tokens. Muse Wonder Full is reserved for Muse Code subscribers only. Both models will also power the Muse app, starting with paying users. CEO Mark Zuckerberg said the models take Meta to the cutting edge of AI.

A new chapter for Meta’s AI team

The Wonder models were built at Meta’s Prometheus data center, a gigawatt-scale facility built for AI training. Meta has spent billions on AI infrastructure and talent over the past year. Its earlier Llama models fell behind rivals from OpenAI, Anthropic, and Google. The Wonder launch shows that investment is paying off.

For developers, the launch adds another strong option. Companies building AI agents can now choose between models from Meta, OpenAI, Anthropic, and Google. For regular users, the new models will first appear inside the Muse app and Muse Code.