TokenPad

Cost scenario · Automation

Classifying support tickets

Every inbound ticket gets routed to a queue, tagged with urgency, and checked for whether it mentions a competitor. One model call per ticket, one short structured answer.

This is the workload where the price difference between model tiers matters most: the task is close to saturated on small models, and the request count is the highest of anything here.

The shape of one ticket

900input tokens
60output tokens
65%of input is cacheable
1model calls per ticket

A 550 token prompt carrying category definitions and two examples, constant across every call, plus a ticket averaging 350 tokens. Output is a small JSON object.

What it costs per month

Computed from the verified rates in the price table, last reviewed August 5, 2026. Figures assume caching is enabled — the uncached column shows what skipping that flag costs.

Monthly cost by volume and model for Classifying support tickets
VolumeMinistral 3 8BCheapest that fitsMagistral MediumMid tierClaude Fable 5Frontier
Small team10,000 tickets / month$1.44$1.44 uncached$21.00$21.00 uncached$67.35$120.00 uncached
Mid-market250,000 tickets / month$36.00$36.00 uncached$525.00$525.00 uncached$1,683.75$3,000.00 uncached
Enterprise3,000,000 tickets / month$432.00$432.00 uncached$6,300.00$6,300.00 uncached$20,205.00$36,000.00 uncached

On this shape the spread between the cheapest model that fits and the frontier one is 83.3×, which is the largest single decision on this page. Enabling prompt caching removes a further 44% on Claude Fable 5, because 65% of the input on this workload is a stable prefix.

Where a provider publishes no cached input rate, the cached and uncached figures above are identical. That is a gap in what the provider discloses, not a model that cannot cache.

What actually drives the bill

Volume, not size

Each request is tiny. What makes this expensive is doing it three million times, which means every token in the constant prompt is multiplied by three million.

The prompt is mostly constant

Category definitions and examples are around 550 of the 900 input tokens and never change. That makes this the most cacheable workload on this site — and the one where forgetting to enable caching costs the most.

Model tier

Classification into defined categories is close to saturated on small models. The price gap between tiers is often more than tenfold, and at this volume that gap is the entire decision.

How to cut it, in order of return

  1. Use the cheapest model that passes your eval setup to 95%

    Not the cheapest model — the cheapest that scores acceptably on your own cases. On classification that is usually the bottom tier, and the difference at three million requests is enormous.

  2. Enable prompt cachingabout 55%

    Two thirds of the input is a constant prefix. Nothing else on this site benefits from caching as much.

  3. Batch anything not needed within the houranother 50%

    Backlog classification, overnight re-tagging and re-processing after a category change are all batch workloads. Only live routing needs a synchronous call.

  4. Skip the model on the obvious cases10–30%

    A password reset request matching a keyword rule does not need a model. Route the unambiguous fraction with rules and send the rest.

What people get wrong budgeting this

  • Adding a third example to the prompt looks free and costs 200 tokens on every one of three million requests.
  • Output tokens are small here, so the usual "output is the expensive side" rule does not apply — this workload is input-dominated and should be optimised as such.
  • Re-classifying historical data after a category change is a batch job. Running it synchronously doubles its cost for no benefit.

Price your own version

The shape above is representative, not measured from your system. Put your real prompt and volume through these and the number becomes yours.

Before committing to volume, the cost control checklist is ordered by saving per hour of effort, and every rate used here links to the provider page it came from on the Mistral AI and other provider pages.