What will this actually cost per month?

How this is calculated

Billable requests are requests × (1 + retry rate), because a call that failed and was retried was still charged, and so was the one that escalated to a human afterwards. Input tokens split by your cache hit rate and are priced at the cached and uncached rates separately. Output tokens are priced whole - caching does not apply to them, which is why output is frequently the line that dominates a chat workload.

Evaluation adds eval set size × runs per month requests at the same token shape. That line is usually a surprise, and it is a cost worth paying: cutting spend without a way to detect the quality you gave up is how a cost problem becomes an accuracy problem you will not notice for weeks.

Cost per resolved task, not per call

Cost per call flatters every system, because the failures are cheap and resolve nothing. The figure that matters divides all the spend - including the calls that escalated - by the tasks actually completed without a human, then adds the human handling for those that were not. That last term is usually the largest and is the one most often omitted.

What this deliberately does not model

Agent loops. If a run can take an unbounded number of steps, a per-request average is not a meaningful unit. Set a hard step ceiling and a per-request cost cap first, then come back to this.

Retrieval and infrastructure. The vector index, hosting and observability are real and are not model spend. They belong in the cost-per-task figure and are specific to your setup.

Batch and off-peak rates, volume commitments, and self-hosted inference. All three can change the answer by a large factor and none of them are general.

The reductions worth making, sorted by how much quality risk each one carries, are in Cutting LLM costs without degrading quality. For the metric this page reports in the fourth box, see cost per resolved task.