18 August 202610 min readDocument processing

Invoice extraction at 100,000 pages a month: fixed cost per page versus per-token pricing

TL;DR

  • At 100,000 invoice pages a month with a 1.5× retry multiplier, GPT-4.1 Mini costs about €210, GPT-4.1 about €1,050 and Claude Sonnet 5 about €1,740 — compute only.
  • Ækora’s Scale plan plus overage is €7,249 at 100,000 pages, or €0.0725 per page, and that price is the finished record — schema, validation, exception queue and delivery — not raw inference.
  • A DIY reserved GPU is about €600 a month before people. Add half an engineer at a loaded €6,000 a month and the line is closer to €3,600.
  • Per-token cost is linear: at 500,000 pages the Sonnet 5 line is 20× the 25,000-page line, while one reserved card runs 500,000 routine pages in under 60% of a month.
  • Under about 25,000 pages a month or spiky volume, stay per-token. Above roughly 50,000 pages against Sonnet 5 or 85,000 against GPT-4.1, with engineers on hand, a reserved GPU wins.

Questions people ask

How much does invoice extraction cost per page?
It depends on the pricing shape. At list prices a routine invoice page costs about €0.0014 on GPT-4.1 Mini, €0.0070 on GPT-4.1 and €0.0116 on Claude Sonnet 5 — before a realistic 1.5× for retries, and compute only. A managed fixed-price pipeline is €0.0725 per page at 100,000 pages a month and includes schema, validation, exceptions and delivery. A reserved GPU is about €0.006 per page in hardware at that volume, before people.
What does it cost to process 100,000 invoice pages a month with GPT-4.1 or Claude?
On list prices at the time of writing, converted at $1 = €0.92, and a routine page profile, 100,000 pages cost about €700 on GPT-4.1 and €1,160 on Claude Sonnet 5 at sticker, or €1,050 and €1,740 with a 1.5× multiplier for retries and second passes. Move 20% of the pile into the hard-page profile and those rise to about €1,460 and €2,390. Those figures are compute only — no prompts, validation, exception review or ERP integration.
Is a fixed-price document pipeline cheaper than paying per token?
Not on the rate card, and the two should not be compared directly. At 100,000 pages a month the managed plan is €7,249 against €1,050 of GPT-4.1 tokens, but the plan price is the finished record — schema, validation rules, an exception queue and CSV, JSON or API delivery — while the token bill excludes the engineer who builds that and the person who clears the queue. Put the same scope on each line before comparing.
How much does a DIY GPU for invoice extraction cost?
A compact document model does about 20 pages a minute on an A100-class card, so 100,000 pages is roughly 83 GPU-hours — €150 of on-demand compute at €1.80 an hour, or €0.0015 per page. Reserved for the month, with power and a slice of ops, the line is about €600. Then add people: half an engineer at a loaded €6,000 a month puts the true DIY line closer to €3,600 than €600.
At what volume does a reserved GPU beat per-token invoice extraction?
Counting people on every line, the rule of thumb is roughly 50,000 pages a month against Claude Sonnet 5 and 85,000 against GPT-4.1, with steady volume and platform engineers on hand. Under about 25,000 pages a month, or with spiky volume, per-token on the cheapest model that passes validation is the better buy — a €600 card at 10,000 pages a month is €0.06 per page. Steady volume without an ML team usually means a managed per-page plan.
Why budget invoice processing as a fixed cost per page?
Because finance leads budget a line — pages expected × rate per page — not tokens. At €0.0725 per page, 100,000 pages is €7,249 and 120,000 is €9,049, and the only variance is volume. With per-token pricing the variance comes from scan quality, retry rates, model choice, vendor price changes and the dollar, none of which show up in a purchase order. Whatever option you choose, express it as cost per page at expected volume with the assumptions written down.

Want this worked out on your documents?

We will price a real sample against OpenAI or Anthropic and tell you whether Bulk document processing or a local model on your hardware is the cheaper first step.