Atlas AI Resources

← Back to AI News

AI products

Meta released Muse Wonder, its first frontier AI model. Here is what outside tests show.

Published October 13, 2026

On October 12, Meta released Muse Wonder and Muse Wonder Full, the first frontier models from its Superintelligence Labs. A frontier model is one meant to compete with the most capable AI systems available. Meta says these match or beat the latest from OpenAI and Anthropic on agent, memory, and coding tasks. An independent testing firm, Artificial Analysis, puts them near the top of its rankings, but not clearly ahead, and Meta’s fuller safety report has not been published yet.

What happened

Muse Wonder is available now to developers through the Meta Model API and in Muse Code, Meta’s coding tool. The stronger Muse Wonder Full is only for Muse Code subscribers and, TechCrunch reports, is not offered through the API. Meta says Wonder will later power its Muse assistant, starting with paid subscribers. The Verge notes Meta has given no timeline.

Developers pay per token, a chunk of text roughly the size of a short word: $4 per million tokens sent in and $20 per million generated. Meta is pitching two features. “Active Memory” means carrying what matters about you across conversations. “Active Focus” means sticking with a hard task, trying other approaches when blocked, and, in Meta’s words, checking in “before taking important actions.”

Meta’s benchmark claims are its own. Some tests it highlights, including the MemoryBench memory test, were built internally. Its model card says the rival-model runs Meta did itself were “best-effort” and may not be optimally tuned for those models.

What outside testing shows

Artificial Analysis scored Muse Wonder 52 and Muse Wonder Full 56 on its intelligence index, up from 48 for Meta’s previous model, Muse Spark 1.3. Wonder Full placed in the top three on six of its ten tests. It found the leads were narrow, and on one key work test Wonder Full was within the margin of error of two Claude models. It also measured cost per task: $4.52 for Muse Wonder, below the Claude models it tested but above GPT-6 Astra, Gemini 4 Argon, and GPT-6.1 Sol. On its test of how often a model makes things up, Wonder Full’s rate fell from 33% to 23%, though general knowledge remains a relative weakness.

Why it matters to a normal person

Another serious contender at the top can mean more pressure on prices and features, and Meta says it wants this technology in front of the billions of people who use its apps. The memory pitch is the part to watch. As Axios notes, agents that remember personal details and act for you also bring new privacy and security problems.

Meta’s own model card recommends extra guardrails, such as a person confirming any action that cannot be undone. That is the right order of things. A model that knows to check in is helpful, but its judgment is not permission. The software around it, and you, should decide what it is allowed to send, buy, or delete.

What it does not mean

It does not mean Meta now has the best AI model. Outside scores put it near the top, and The Verge says it is unclear how it compares with Google’s Gemini 4 Argon. It does not mean the models are open for anyone to download. Axios reports Meta is not saying whether it will release them that way.

And it does not mean the safety questions are settled. Meta’s model card says that without safeguards it cannot rule out the model reaching its “high” risk level for chemical and biological misuse, so it added safety training and automated filters, and rates the remaining risk as “moderate or lower.” That is Meta grading its own work, and the fuller report is still to come.

Sources