Muse Wonder and Muse Wonder Full: what Meta's new AI models can do

What Meta has introduced

Muse Wonder is a new model series from Meta Superintelligence Labs, the AI division Mark Zuckerberg reorganised a little over a year ago and placed under former Scale AI chief Alexandr Wang. A model is the program in the background that understands language, plans and produces answers. Products such as the agent app Muse or the programming tool Muse Code are built on top of it.

The series has two variants. Muse Wonder is the standard model. According to Meta, Muse Wonder Full is the more powerful version and Meta's most capable model so far. Both process text, images, video, audio and PDF documents but only answer in text. They can take very long material into account at once, around one million tokens, roughly a few thousand pages of text.

The predecessor is called Muse Spark. Meta introduced it in April 2026 and developed it up to version 1.3. It also powers the personal agent Muse, which Meta presented in September. According to Mark Zuckerberg, Wonder is a much larger and more advanced class of models. It was trained in Meta's Prometheus data centre.

Unlike the earlier Llama series, Meta does not make the Wonder models available for download. They can only be used through Meta's own services. Meta offers a small open model called Muse Glimmer separately.

Active Memory and Active Focus: the two new abilities

Meta puts two properties at the centre that matter above all for agents, meaning AI that completes tasks by itself.

Active Memory: the model is meant to remember across several conversations what matters to you and bring earlier details back on its own when they count for a new task. If you once told an agent that you go running on Saturdays and don't eat meat, you should not have to repeat that for every plan.

Active Focus: the model is meant to stick with a goal when things get difficult. Instead of stopping at an obstacle, it tries other approaches, fixes errors and asks when something is unclear. Before important steps it is supposed to check in. Alexandr Wang told CNBC the model was trained specifically for this because agents had so far failed at everyday tasks such as table reservations.

As examples, Meta mentions handling an insurance claim, understanding a lease, sorting out personal finances or programming a small app. In demonstrations, Muse Wonder plans a running route based on personal preferences, designs merchandise from Instagram photos and creates a seating chart for a wedding. Whether this works as well in everyday life as in a demo cannot be said yet: the app with Wonder has not been released to users.

How good are the models? The independent numbers

The most meaningful figure is the Intelligence Index from Artificial Analysis, a testing firm that independently compares AI models with ten tests covering reasoning, knowledge, coding and office work. Muse Wonder Full reaches 55.8 points there, 56 when rounded. That ties Claude Sonnet 5.5 and is only behind Claude Opus 5.5 at 57.6. Behind it come Claude Fable 5.1 at 53.4, GPT-6 Astra at 52.7 and Google's Gemini 4 Argon at 52.6. The standard Muse Wonder scores 52.4. Its predecessor Muse Spark 1.3 stood at 48.1, only six weeks before Wonder's launch.

The biggest progress shows in independent work. In a test where the AI has to solve tasks on a computer via the command line, Muse Wonder Full rose from 33 to 57 percent according to Artificial Analysis. In agentic programming, it ties Anthropic's Claude Code with Claude Sonnet 5.5 when running in Meta's own tool Muse Code. It also makes up facts less often: according to Artificial Analysis, its rate of such hallucinations fell from 33 to 23 percent.

The testing service Vals AI places Muse Wonder Full fifth in its ranking, ahead of all OpenAI models. Only current Anthropic models rank higher there.

One caveat: the gaps at the top are small. On some individual tests, Wonder Full is only narrowly ahead or within the statistical margin of error, according to Artificial Analysis.

Meta's own benchmarks: strong on long programming tasks

Meta itself has published a longer list of test results. They come partly from third parties, partly from Meta's own runs and partly from figures supplied by the other providers. Some programming tests were run with different tools. Such manufacturer figures are therefore less strictly comparable than independent tests.

Muse Wonder Full leads most clearly on SWE-Marathon, a test for long, multi-part software projects. It reaches 57.5 percent there, Claude Opus 5.5 scores 46.3 and GPT-6 Astra 42.5. On a legal work test by the law-firm software company Harvey, Wonder Full reaches 13.1 points, Opus 5.5 only 4.2. The low values also show how far all models still are from error-free legal documents. On the business workflow test AutomationBench, Wonder Full also leads with 45.4.

It does not lead everywhere: on health questions in HealthBench Pro, on command-line tasks in Terminal-Bench 4.0 and on office tasks in JobBench, Claude Opus 5.5 or GPT-6 Astra are level or ahead.

Wonder's clearest weakness is pure factual knowledge. In Artificial Analysis's AA-Omniscience knowledge test, Wonder Full scores only 30 points, Wonder 25. Claude Opus 5.5 scores 46, GPT-6 Astra 43. The specialist site The Decoder suspects a smaller base model than the competition's. For you, this means: for facts from history, science or general knowledge, checking is still worthwhile.

What it costs and who can use it

Developers can use Muse Wonder through Meta's programming interface. One million input tokens cost 4 US dollars, one million output tokens 20 US dollars. That matches the prices of Claude Opus 5.5 exactly and is cheaper than GPT-6 Astra at 10 and 50 US dollars. The predecessor Muse Spark 1.3 cost only 1.25 and 4.25 US dollars.

What matters, however, is the cost of a completed task, because models use different numbers of tokens. According to Artificial Analysis, an average test task costs 4.52 US dollars with Muse Wonder, almost three times as much as with Muse Spark 1.3 (1.60). Claude Opus 5.5 costs 5.98, GPT-6 Astra 3.26 and Gemini 4 Argon 1.99 US dollars. One reason: Muse Wonder reads a great deal of text per task.

Muse Wonder Full is not yet offered through the API, only in the Muse Code programming subscription for 50 or 200 US dollars a month. Important for privacy: according to VentureBeat, these subscriptions default to the so-called Contributor mode, in which Meta may use submitted code, files and responses for training. You can switch this off, but the change only applies to future submissions. Through the paid API, Meta says content is not used for training.

Wonder is due to be used in the agent app Muse soon, starting with subscribers. According to CNBC, these subscriptions cost 20 or 100 US dollars a month. Meta has not given an exact date. For Germany and the rest of the EU, little changes for now: Meta does not offer the Muse app in Europe and has given no date for it.

Our conclusion: with Wonder, Meta has reached the top of AI development without leading it. The interesting question is whether the better model fixes the known problems of the Muse app, such as made-up details or abandoned bookings.