Meta's Muse Wonder Full Ranks Top-Tier but Costs Nearly 3x More

What happened

Meta has released a new frontier AI model, Muse Wonder Full, which now ranks among the top three performers on six out of ten industry benchmarks, according to The Decoder. The model shows its strongest gains on agentic tasks — benchmarks that test an AI system's ability to plan, use tools and complete multi-step actions rather than simply answer questions.

The improved capability comes at a cost: running Muse Wonder Full is reported to be nearly three times more expensive per task than its predecessor. The Decoder's reporting positions this as part of a broader pattern among frontier labs, where each generational leap in capability is increasingly paired with a steep rise in compute cost per query.

Why it matters

Muse Wonder Full's benchmark position matters less for the ranking itself than for what it signals about where frontier AI is heading: toward models built explicitly for agentic work — executing tasks, not just generating text. For organisations building AI into operations, customer service or internal workflows, this is the capability class that actually changes what automation can be trusted to do unsupervised.

But the near-tripling of cost per task reframes the adoption conversation. Capability gains are no longer "free" improvements layered onto existing pricing; they come with a direct cost-to-serve trade-off. Any organisation evaluating frontier models for production use now has to weigh marginal performance gains against materially higher unit economics — a calculation that will shape which use cases justify top-tier models versus smaller, cheaper ones.

By the numbers

• Top three ranking achieved on six of ten benchmarks evaluated.
• Nearly tripled cost per task compared with the prior model generation.

The Renascence take

The headline framing — "breaks into the AI top tier" — is the least interesting part of this story. The real signal is the cost curve, and it's one that experience and transformation leaders consistently underweight when they get excited about a new model's benchmark scores.

Benchmark leaderboards measure capability in a vacuum; they say nothing about whether a capability is affordable to deploy at the scale a real operation requires. A model that is marginally better at agentic tasks but nearly three times costlier per task doesn't automatically win the business case — it shifts the question from "can this model do the job" to "which jobs are valuable enough to justify this model doing it." Operators chasing frontier capability should be building a tiered deployment strategy now: premium models for the handful of high-stakes, high-complexity interactions where the margin justifies the spend, and leaner models for the long tail of routine tasks. The organisations that get burned by frontier AI won't be the ones that adopted it too slowly — they'll be the ones that deployed the most expensive model everywhere, by default, without doing this arithmetic first.