Cost scenario · Automation
Processing documents at scale
Invoices, contracts or reports arriving continuously. Each one is read, key fields are extracted into structured data, and a short summary is produced for a human reviewer.
Documents are long and answers are short, which inverts the usual advice: here the input side is the bill and output optimisation barely moves it.
The shape of one document
A document averaging 20,000 tokens — roughly 15 pages — plus a 1,000 token extraction prompt with the schema, and around 1,000 tokens of formatting instruction. Output is a JSON object plus a paragraph.
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.
| Volume | Ministral 3 8BCheapest that fits | Magistral MediumMid tier | Claude Fable 5Frontier |
|---|---|---|---|
| Small operation2,000 documents / month | $6.81$6.81 uncached | $95.00$95.00 uncached | $490.20$510.00 uncached |
| Established50,000 documents / month | $170.25$170.25 uncached | $2,375.00$2,375.00 uncached | $12,255.00$12,750.00 uncached |
| High volume500,000 documents / month | $1,702.50$1,702.50 uncached | $23,750.00$23,750.00 uncached | $122,550.00$127,500.00 uncached |
On this shape the spread between the cheapest model that fits and the frontier one is 74.9×, which is the largest single decision on this page. Enabling prompt caching removes a further 4% on Claude Fable 5, because 5% 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
The document is 90% of the request
Everything else is rounding. Halving the prompt saves five percent; halving the document you send saves forty-five.
Almost nothing caches
Every document is different, so only the schema and instructions cache — about five percent of input. This is the workload where prompt caching helps least.
How to cut it, in order of return
- Send only the pages that matter40–70%
Most extraction targets live on specific pages. A cheap classifier or a layout heuristic that identifies them first removes most of the input, and it is the single largest saving available here.
- Batch the whole pipelineabout 50%
Document processing is almost never interactive. If a result is acceptable within a few hours, batch pricing halves the bill for a configuration change.
- Strip boilerplate before sending10–25%
Headers, footers, legal boilerplate and page numbers repeat on every page and carry no extractable information.
- Use a cheaper model for extraction, expensive only for summary30–50%
Pulling a named field from text is an easy task. Producing a good summary of a contract is not. Splitting them lets each run on the right tier.
What people get wrong budgeting this
- Scanned documents need OCR first, and OCR errors become extraction errors that look like model failures.
- Chunking a document that nearly fits is worse than sending it whole — the overlap is pure duplication.
- Re-processing after a schema change means paying for the whole corpus again. Store the raw extraction, not only the mapped result.
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.
- Document Token ProfilerSection-by-section token share, with a density flag for dense content.
- Text Splitter for RAG and EmbeddingsToken-accurate chunking with configurable overlap. Respects sentence and paragraph boundaries.
- Batch API Savings CalculatorThe 50% discount most teams never claim, priced at your volume.
- LLM API Cost CalculatorRequests per month in, dollars out. Input, cached input and output priced separately.
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.