Meta's Muse Wonder Full Ranks Top 3 on 6 of 10 Benchmarks, but Cost Per Task Nearly Triples

Meta is back in the conversation about the world's best models. According to The Decoder's summary, the new Muse Wonder models place in the top three on six of ten benchmarks. The most noticeable improvement over Meta's previous results comes in agentic tasks, meaning scenarios where the model builds its own plan, calls tools, and carries multistep work through to a result. That came at a price: the cost of completing a single task has risen almost threefold.

What the tests showed

Six top-3 placements out of ten possible mean the model has become competitive across a broad front rather than winning in one narrow discipline. It also means Muse Wonder Full did not make the top three on four benchmarks. The available summary does not say which tests those are, and that is the first thing to check in primary sources.

Key figures known so far:

• Result: top 3 on 6 of 10 benchmarks.
• Strongest area: agentic tasks, where the gain is largest.
• Cost: the price of completing a single task has risen almost 3x.
• Lineup: several Muse Wonder models; the flagship version is called Full.
• Not disclosed in the summary: model size, context window, per-token pricing, access terms, and the list of benchmarks.

Why "cost per task" matters more than "cost per token"

Cost per task is calculated differently from a per-million-token rate. It combines the rate, the number of reasoning tokens, the number of tool calls, retries, and the length of the final chain of actions. So a nearly threefold increase does not necessarily mean Meta tripled its price list. The model may simply spend more steps and tokens on each task. That is an assumption, not an established fact: the summary contains no cost breakdown.

This trade-off is typical for agentic scenarios. The longer a model thinks and the more it checks its own actions, the better its chance of finishing the task, and the higher the bill. The question is whether the quality gain pays for itself. If the model solves noticeably more tasks on the first attempt, the cost of each successfully completed task may rise far less than the cost of an attempt. If the benchmark gain is only a few percentage points, a threefold increase in spending looks hard to justify.

What to check before adopting it

Until the full data is published, it is wisest to treat the results as a bid for a place among the top tier, one that needs confirming on your own tasks. A practical sequence:

• Take 30–50 real work tasks, not a public benchmark, and run them through Muse Wonder Full and your current model.
• Measure the cost of a successfully solved task, not the cost of a request. Include retries and tool calls.
• Find the four benchmarks where the model missed the top three and compare them with your workload profile. If they cover coding or long context and that is exactly what you work on, the conclusions will be different.
• Clarify access terms: which API, what limits, and in which region the model is available.
• For high-volume, simple scenarios, keep the cheaper model, and reserve Muse Wonder Full for complex agentic chains where mistakes cost more.

The main takeaway for now is simple: Meta once again has a model that cannot be ignored when choosing a frontier system. But whether it is worth three times the price will have to be decided from numbers that are not yet publicly available.