# Meta has released Muse Wonder and says it can compete with the best AI models. How good is it? And which version can you actually use?

Source: https://maito.app/blog/muse-wonder-ai-models
Date: 2026-10-12

Two new models, promising agent benchmarks, and an important difference between the version you can call and the one leading Meta’s charts.

## What did Zuckerberg just announce?

Two new AI models: **Muse Wonder** and **Muse Wonder Full**. Meta presents them as its strongest models yet, designed for coding, research, document work, and personal tasks that require several steps. Zuckerberg’s announcement includes a chart comparing them with leading models from Anthropic and OpenAI. [Announcement](https://x.com/finkd/status/2109689305073614918).

The launch follows Meta’s overhaul of its AI operation. The company says Wonder builds on the foundation established by Muse Spark, its earlier model family. The ambition is to create assistants that can work through substantial assignments while keeping track of your requirements. [Meta’s launch announcement](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/).

## Everyone keeps saying “agentic.” What would that actually look like?

Think of asking an assistant to build a small application. Completing the assignment could involve inspecting existing files, planning changes, writing code, running tests, discovering a problem, and trying another approach.

An agent combines a model with tools and software that lets it carry out that sequence. The model helps decide what to do; the surrounding system executes actions and returns their results.

Meta positions Wonder for these longer workflows, including tasks split among several agents. Its ready-made coding agent is **Muse Code**; developers can also use the model through an API to build their own applications. [Meta’s developer documentation](https://dev.meta.ai/docs/overview).

## Fine, but is it actually competitive with Claude?

There is independent evidence that **Wonder Full** is competitive.

Artificial Analysis currently reports these scores on its Intelligence Index:

| Model and evaluated configuration                        | Score |
| -------------------------------------------------------- | ----: |
| Muse Wonder, maximum reasoning                           |    52 |
| Muse Wonder Full, maximum reasoning                      |    56 |
| Claude Opus 5.5, maximum reasoning with default fallback |    58 |

Sources: [Muse Wonder](https://artificialanalysis.ai/models/muse-wonder), [Wonder Full](https://artificialanalysis.ai/models/muse-wonder-full), [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5).

That supports a narrower conclusion than “Meta has the best model”: its stronger version performs close to a leading competitor on this particular evaluation.

The index combines ten tests covering areas such as agent work, coding, knowledge, and scientific reasoning. It is primarily an English-language, text-based evaluation. It cannot establish how well a model handles every task you might give it. [Evaluation methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking).

## But Zuckerberg’s chart shows Wonder beating other models. What’s the catch?

The results vary by task. Here are three coding rows from Meta’s chart:

| Benchmark                                      | Wonder Full | Claude Opus 5.5 |
| ---------------------------------------------- | ----------: | --------------: |
| DeepSWE 1.1: work across software repositories |        79.0 |            74.2 |
| Terminal-Bench 4.0-AA: tasks in a terminal     |        57.1 |            59.6 |
| ProgramBench: reconstructing complete programs |        59.0 |            70.5 |

Wonder Full leads the repository benchmark in this comparison, while Opus leads the other two. These are different evaluations; their scores should not be averaged into a homemade overall ranking. [Meta’s published chart](https://x.com/finkd/status/2109689305073614918).

There are also differences in how results were obtained. Meta’s evaluation report combines third-party results with internal testing and provider-reported figures. Some coding comparisons use different agent software. The report acknowledges that internally tested competitors may not have optimally tuned prompts or tools. [Meta’s evaluation report](https://research.meta.ai/static/muse-wonder-model-card).

The chart is useful evidence about particular strengths. It does not settle which model will perform best inside your preferred coding tool.

## What are “Active Memory” and “Active Focus”? More marketing names?

They are Meta’s names for two useful goals.

**Active Memory** means carrying relevant information across conversations and retrieving it when it becomes useful. For example, an assistant helping with an event should remember that two guests cannot sit together when it later prepares the seating plan.

**Active Focus** means continuing toward a difficult goal: trying alternatives, adapting when circumstances change, and asking for clarification when necessary. [Meta’s research explanation](https://research.meta.ai/blog/muse-wonder).

The distinction matters. An assistant can remember your requirements and still fail to complete the assignment. It can also work persistently while pursuing the wrong interpretation.

Wonder’s roughly one-million-token context window gives it room to process substantial material in a request. That capacity alone does not prove reliable memory across future conversations. [Model specifications](https://artificialanalysis.ai/models/muse-wonder).

## Can I try the stronger model through the API?

**No. Wonder Full is currently restricted to Muse Code subscriptions.**

The standard Wonder model, identified as `muse-wonder-1.0`, is available through the Meta Model API and Muse Code. Full requires the **Power Usage** or **Maximum Usage** subscription.

Setting standard Wonder’s reasoning effort to maximum does **not** turn it into Full. They are separate variants. [Availability documentation](https://dev.meta.ai/docs/models).

Meta’s launch page lists the eligible plans at **$50 and $200 per month**, respectively. [Plans and pricing](https://dev.meta.ai/models/muse-wonder).

The consumer Muse assistant is another surface: Meta says Wonder will roll out there first to subscribers. That is a rollout announcement, rather than confirmation that everyone using Muse already has it. [Launch announcement](https://about.fb.com/news/2026/10/introducing-muse-wonder-models/).

## And what does the standard model cost?

Meta’s API prices are:

| Token category | Price per million tokens |
| -------------- | -----------------------: |
| Input          |                       $4 |
| Cached input   |                    $0.20 |
| Output         |                      $20 |

Tokens are the pieces of text a model processes and generates. Cached input is previously processed material that qualifies for the lower rate. [Official pricing](https://dev.meta.ai/docs/pricing-rate-limits).

For illustration, a request with 100,000 uncached input tokens and 10,000 output tokens would cost **$0.60 in token charges**. That is a calculation from the listed rates, not a measured cost for completing a real project.

An agent can make many requests while attempting one assignment. The useful comparison is therefore how much a completed task costs—including retries and corrections.

## This is Meta. Can I download the weights?

**Wonder and Wonder Full are proprietary models.** Their weights are not publicly available, according to Artificial Analysis’s listings. [Wonder](https://artificialanalysis.ai/models/muse-wonder), [Wonder Full](https://artificialanalysis.ai/models/muse-wonder-full).

Meta does offer an open-weight model called **Muse Glimmer**, which developers can download and run on their own hardware. It is a separate model, distilled from Muse Spark. Downloading Glimmer does not give you a locally hosted version of Wonder. [Meta’s model overview](https://dev.meta.ai/docs/overview).

For anyone whose priority is local deployment or control over the model itself, that distinction is consequential.

## If I use Muse Code, will Meta train on my work?

**That depends on the mode—and subscriptions default to Contributor mode.**

In Contributor mode, subscribers receive more usage in exchange for allowing Meta to use submitted content to train future models. The content includes prompts, code, files, context, and responses.

You can disable Contributor mode in settings. Meta says future content then will not be used to train its models unless you enable it again. [Subscription documentation](https://dev.meta.ai/docs/muse-code/subscriptions).

Standard Wonder access through the pay-as-you-go API uses Standard mode, where Meta says content is not used for training. [Model documentation](https://dev.meta.ai/docs/models).

This is an access and data-use choice worth checking before bringing a private repository into the workflow.

## So should I switch from the model I already use?

The evidence makes Wonder worth evaluating for demanding agent and coding work. It does not establish that switching will improve your particular workflow.

Start with an assignment you understand well: a repository change with verifiable tests, or a document task whose sources you can check. Compare the finished result, elapsed time, total cost, and how often you have to intervene.

Test the version you can actually access. The standard API model’s Intelligence Index score is 52; the stronger, subscription-only Full version scores 56. A review of Full cannot tell you exactly what to expect from standard Wonder. [Independent evaluations](https://artificialanalysis.ai/models/muse-wonder).

The decisive result is whether it completes your assignment accurately with less supervision. A model that wins your task earns the switch.

_Details checked October 12, 2026, at 19:29 UTC._