Meta’s Muse Wonder frontier claim: what to know

Artificial Analysis scores Muse Wonder Full at 56 and Muse Wonder at 52 on its Intelligence Index. The index combines 10 evaluations spanning areas such as agentic knowledge work, business tasks, coding, science and reasoning. It offers a comparison across that test suite, not a verdict on every task a model might be asked to do.

Artificial Analysis lists Claude Opus 5.5 at 58 on the same index, two points above Wonder Full. That result supports a claim of competitiveness on this particular measure; it does not establish that Meta has matched leading models across uses or tasks.

What can Muse Wonder do?

Meta says the models build on Muse Spark and work in pretraining and reinforcement learning, methods used to train and refine models. The company also describes Active Memory as retrieving useful information across conversations and Active Focus as continuing toward a goal, adapting when obstacles arise and seeking clarification.

Those features are intended to support longer-running agent tasks, but the descriptions do not show how consistently the models handle them in ordinary use. The supplied reporting contains no independent tests of Meta’s specific claims about legal, financial, health or coding work, or evidence from users trying extended workflows.

Who can use Muse Wonder now?

Meta says Muse Wonder is available through its Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers, rather than through the API, based on the access routes described at launch. Engadget reported that Full access is included in Muse Code’s $50-a-month Power Usage and $200-a-month Maximum Usage plans.

That split matters to developers weighing which model to test or integrate: Wonder has an API route, while the reporting does not establish API access to Full. The available information also does not include user or developer accounts of how either model performs in practice.

Meta says Wonder will later power experiences in its Muse assistant, first for subscribers, but has not announced a date. That planned rollout is not current availability in the assistant.

What evidence is still missing?

An index score can help compare models on a consistent set of tests, but it cannot establish whether a model is dependable in a particular workplace or extended task. Independent evaluations of the legal, financial, health, memory and coding uses Meta highlights would provide evidence beyond the overall score.

Safety is a separate question from benchmark performance. The Intelligence Index compares capabilities on its evaluations; it is not a safety assessment. Meta says it trained safeguards and plans to publish a fuller safety-evaluation report later, but that report was not available with the launch materials.

For people choosing or building with these models, the current evidence establishes a benchmark comparison and distinct access options—not broad parity or proven reliability across real-world uses. Meta’s assistant rollout and fuller safety report remain outstanding.