Meta has released two new models in the top performance class: Muse Wonder and Muse Wonder Full. According to the company, both are taking on the current top models from OpenAI and Anthropic. Frontier models are those systems that mark the current state of performance and against which the other providers are measured.

In benchmarks — standardised test series used to evaluate models across comparable tasks — the new models reach a place among the top three in six out of ten cases. That is not enough for a consistent leading position, but it puts Meta into the narrower top group.

Clear gains in agentic tasks

The models show the greatest progress in agentic tasks. This refers to processes in which a model not only produces a single answer but plans several steps, calls tools, checks intermediate results and continues working on that basis. This is precisely the type of task that is relevant for development and automation projects: research across multiple sources, data reconciliation between systems, code changes including testing and correction.

The fact that a larger leap is reported in precisely this area fits the general direction of development in recent months. Providers are increasingly optimising less for individual answers and more for reliability across long task chains. That is where it is decided whether a model can be used productively or whether the effort for monitoring and rework consumes the benefit.

The price of progress

The increase in performance does not come for free — literally and figuratively. According to the available information, the costs per task have almost tripled. That is a value that cannot readily be translated into token prices: agentic models work through more intermediate steps, generate longer chains of thought and call tools more frequently. So the bill rises not only because a token becomes more expensive, but simply because more work is involved per task.

There is also the observation that Meta's new models are no longer available free of charge. The company had previously pursued a comparatively open course. The switch to a paid offering changes the starting position for everyone who has so far had Meta models in their toolbox as a low-cost alternative. There is no reliable information on specific prices, pricing models or usage limits; here it is worth checking the official terms before planning budgets.

What the two sources emphasise differently

The available reports set different priorities. One emphasises the benchmark success and the leap in agentic tasks together with the increased costs per task, the other foregrounds the loss of free use and places the models in the competition with OpenAI and Anthropic. The two accounts do not contradict each other but describe two sides of the same development: more performance, higher running costs.

How robust the benchmark results are in everyday use cannot be judged from a distance. Test series reflect defined task formats and say little about how a model copes with your own data, your own domain language and your own interfaces. A place among the top three is an indication, not proof.

Consequences for tool selection

For projects that firmly plan for AI support, the basis for decisions is shifting. Previously the question was usually: Which model solves the task best? In future it will be more like: Which model solves the task well enough, at what cost per run?

  • Separate tasks. Not every work step needs a frontier model. Classification, summarisation or format conversion can often be handled with smaller, significantly cheaper models. The expensive variant remains reserved for multi-step, agentic tasks.
  • Measure costs per task, not per token. For budget planning, the complete chain is decisive — including repetitions, failed tool calls and corrections.
  • Ensure interchangeability. Anyone who encapsulates model access behind their own abstraction layer can switch providers without rebuilding the application. Given the current pace of releases, this is no theoretical exercise.
  • Check licence and access conditions. A model that was freely usable yesterday may be licensed differently today. This affects costing, data protection and contract design alike.

Assessment for practice

Competition at the top means that agentic capabilities improve faster than many project plans can keep up with. For web and software projects, this means: model selection should be an interchangeable component of the architecture and not a commitment that cannot be revised for years.

At the same time, the question of the cost per completed task belongs in the profitability calculation from the outset — a more powerful model only pays off where the manual effort saved actually exceeds the higher bill.

Sources