Third-party tests found that Meta’s Muse Wonder models improved on the company’s previous Muse Spark 1.3, but the gains were not a clean sweep: the higher-scoring Wonder Full led only some of the tested tasks and is currently limited to a Muse Code subscription.
Illustrative photo by PXHERE, dedicated under CC0 1.0 via Wikimedia Commons.
Scores show gains across a mixed test set
Artificial Analysis gave Muse Wonder Full 56 and Muse Wonder 52 on its Intelligence Index v4.3, compared with 48 for Muse Spark 1.3. The scores were measured at maximum reasoning effort. Wonder Full placed in the top three on six of the index’s ten evaluations, including narrow leads on CritPt and GDP.pdf. These are benchmark results, not a measure of every kind of work an agent may face.
Some results were weaker. On AA-Omniscience, which tests knowledge and hallucination, Wonder Full scored 30, level with Muse Spark 1.3 and below several competing models. Artificial Analysis described knowledge as a relative weakness. The pattern points to meaningful gains in selected tests, not a general lead across capabilities.
The strongest model has narrower access
Meta lists Muse Wonder for both its Model API and Muse Code. The API is priced at $4 per million input tokens and $20 per million output tokens, with a lower rate for cached input. Muse Wonder Full, which received the higher Artificial Analysis score, is currently available only to Muse Code subscribers. Meta says Wonder will eventually power more features in its Muse assistant, beginning with subscribers to Muse.
Memory claims still need real-world evidence
Meta says Active Memory can carry relevant context between conversations, while Active Focus helps an agent continue after a blocker, try another route and ask before important actions. Its model card describes an internal MemoryBench test built around long conversations, distracting details and later corrections. That test does not establish how reliably memory works in everyday, long-running use outside the lab.
The same card says the models can make errors in multi-step reasoning, especially in novel situations, and that results vary with prompts, tools and the surrounding agent software. Meta recommends added safeguards such as human confirmation before irreversible actions.