Meta released Muse Wonder Full - and that's where it gets interesting
Meta introduced two new models: Muse Wonder and the more powerful Muse Wonder Full. According to Artificial Analysis, Full scored 56 points and entered the top-3 on six out of ten tests. Growth is particularly noticeable in agentic tasks: on Terminal-Bench, the result increased from 33% to 57% compared to Muse Spark 1.3.
But there's a nuance that's easy to miss behind the beautiful benchmark.
Performance increased - and so did cost. For Muse Wonder, the estimated task price is $4.52 versus $1.60 for Spark 1.3 - nearly three times more expensive. Meanwhile, there's no public API pricing for Wonder Full yet: the model is only available to Muse Code subscribers.
For developers of agentic systems, the conclusion is simple: the new model should be evaluated not by its ranking position, but on your own scenarios. How many steps does it take? How many tokens does it consume? And how much will a month of operation cost after scaling?
I would start with a small A/B test on real tasks, not with migrating the entire pipeline. Top-3 in the benchmark is a good reason to try the model. But not yet a reason to give it your entire budget.
Sources: The Decoder , Latent Space