Meta shipped Muse Wonder and Muse Wonder Full this week, calling them its most capable models. Both are 1M-context models pitched at agent work. Muse Wonder runs $4/$20 per million tokens on the Meta Model API. Wonder Full isn’t on that menu: it lives on Muse Code plans, where it reportedly leads DeepSWE and SWE-Marathon.
Artificial Analysis puts Muse Wonder at Max reasoning level on par with GPT-6.1 Sol. That’s the number everyone will quote, and it’s the one that needs an asterisk, because Max is one of six settings.
The dial is the product
Minimal, low, medium, high, xhigh, max. Simon Willison’s llm-meta-ai 0.2 exposes all six, and his pelican-on-a-bicycle SVGs across the range are the most honest artifact to come out of this launch so far: one prompt, one model, six very different answers. Not six quality tiers. Six different cost structures.
This is where the frontier-parity claim gets slippery. If Max matches GPT-6.1 Sol, fine, but Max burns the most reasoning tokens, and reasoning tokens bill at the $20 output rate. The posted price is a floor, not an estimate. Anyone building an agent loop on this needs to figure out which rung of the ladder their task actually needs, and nobody has good tooling for that decision. You guess, you measure, you guess again. The vendors ship the dial and hand you the problem.
I’d rather have the dial than not. The alternative is a single opaque setting tuned for leaderboard screenshots. But “Meta’s new model matches GPT-6.1” and “Meta’s new model matches GPT-6.1 if you pay for maximum reasoning on every call” are different sentences, and only one of them is going on a slide.
Wonder Full in the wrong place
The split between the two models bothers me more than the pricing. Wonder Full is the better coder, the one topping DeepSWE and SWE-Marathon, and it’s behind a Muse Code subscription rather than the API. So the strongest model is reachable by Meta’s own coding product and not by whatever you were planning to build.
OpenAI and Anthropic have both played versions of this game. It works, commercially. It also means the SWE-bench-style numbers that get reported for Wonder Full describe a thing most developers can’t call programmatically. Benchmarks on a subscription-gated model are marketing for the subscription.
What I want to see over the next few weeks: whether Muse Wonder at medium or high holds up on real agent traces, not pelicans. Long-horizon tool use is where 1M context either earns its keep or quietly degrades around turn forty. Context window size has stopped predicting context window usefulness, and no public benchmark has caught up to that gap.
llm-meta-ai 0.2 is on PyPI now, so the six-level comparison is a pip install away.