The market for so-called agentic AI models – that is, models that independently plan multi-step tasks, call tools and check intermediate results – is currently moving in two directions at once. On one side is price, on the other peak performance. Two recent reports illustrate both poles quite clearly.

Meta increases the pace

With Muse Spark 1.3, Meta has released the fourth model in the series within five months. According to the evaluation by Artificial Analysis, a service that compares language models by performance and cost, the model makes significant gains above all in agentic benchmarks. However, that is not enough to reach the top: Claude Fable 5.1 and other top models remain ahead in the assessment.

The strongest argument lies elsewhere. At 0.55 dollars per task, according to this evaluation, Muse Spark 1.3 undercuts all comparably rated competitors. The cost figure does not refer to individual tokens but to the complete processing of a task – a metric that is more meaningful for agentic systems because an agent can generate dozens of model calls for a single request.

The quality peak remains occupied

How far capabilities at the top reach is illustrated by a second report: Claude Fable 5.1 from Anthropic has apparently solved a centuries-old number puzzle that was considered unsolved among researchers – without human help. Details on the solution path and on independent verification of the result are not available in the information provided, so the assessment should be made with the appropriate caution.

For practice, the case is still instructive. It describes exactly the type of task in which top models set themselves apart from cheaper alternatives: long chains of hypothesis, verification and correction, in which a single error early in the process invalidates the entire result. Such tasks are rare, but when they occur, a cheaper price per task does not help.

Why price per task changes the architecture

In classic chat applications, the model price hardly matters because one call is made per user request. Agentic systems reverse this ratio. An agent that processes a ticket, gathers data from several systems and validates a result produces a chain of calls – and each one costs money.

This shifts the calculation. In a mass process with many thousands of transactions per month, the price per task determines whether a use case is viable at all. In a process with few but consequential decisions, the error rate is the more relevant variable because every manual rework is more expensive than the model call itself.

What follows from this for model selection

No general recommendation can be derived from the two reports, but a way of working can:

  • Separate tasks instead of standardising models. Routine steps such as classification, extraction or formatting can often be handled with a cheaper model. The more expensive call remains reserved for the steps where precision matters.
  • Measure costs per transaction, not per token. Only the total number of calls per business case shows what a use case costs in operation.
  • Plan for replaceability. Four model versions in five months are an indication of how quickly price-performance ratios shift. An application that is firmly tied to one provider cannot benefit from this.
  • Build your own test cases. Public benchmarks measure general capabilities. Whether a model can cope with your documents, your specialist language and your special cases is shown only by your own test set of real transactions.

Limits of comparability

The benchmark values and cost figures here come from an external evaluation and refer to a specific task profile. They say little about how a model behaves on tasks that do not appear in the benchmark – for example, with strongly domain-specific content or with requirements for data storage and the legal framework.

Likewise, a price per task does not include the ancillary costs: safeguarding, logging, human review loops, repeat runs on failed attempts. In projects, it is precisely these items that often account for the larger part of the effort.

Assessment

For web and software projects, this above all means: model selection is no longer a one-off decision but an operational variable that is reviewed regularly. Anyone who places an abstraction layer between application and model and measures costs and error rates per transaction can react to price movements without rebuilding the application. The more exciting question is anyway not which model is currently ahead, but which process step needs a top model at all.

Sources