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.
At 100,000 invoice pages a month, the token bill is the small number and the pipeline around it is the large one. On list prices at the time of writing, and with a realistic 1.5× multiplier for retries and second passes, GPT-4.1 Mini costs about €210 for the month, GPT-4.1 about €1,050, and Claude Sonnet 5 about €1,740. A managed fixed-price pipeline — Ækora’s Scale plan plus overage — is €7,249, or €0.0725 per page, and that figure includes the schema, validation, exception handling and delivery that a token invoice leaves out. A do-it-yourself reserved GPU is about €600 a month before people. Which one is cheapest depends on who builds and runs everything that is not the model call.
Everything below is an illustrative scenario with the assumptions stated in each table caption. It is a budgeting model you can rerun on your own token export, not a benchmark and not a claim about anyone’s production numbers.
The three ways invoice extraction gets priced
Every AI invoice processing offer you will see is one of three shapes. They scale differently, and the shape matters more than the headline rate.
Per token. OpenAI, Anthropic, Google and Mistral bill input and output tokens. Cost is pages × tokens per page × retries. It is zero at zero volume, has no ceiling, and moves with scan resolution, prompt length and the vendor’s price list. You still build and run everything around the call.
Per page, as a service. IDP vendors and managed pipelines sell a monthly allowance with an overage rate. Cost is a step function: flat inside the allowance, linear above it. The price usually covers the finished record — extraction, validation, delivery — not only inference.
Fixed plan or reserved capacity. A rented GPU by the month, or an owned box, at a flat cost up to a capacity ceiling. Per-page cost falls as volume rises and is the same whether the card is busy or idle. You own the model, the queue and the on-call.
Per-token pricing is the cheapest way to start and the hardest to forecast. Reserved capacity is the cheapest per page at high, steady volume and the most expensive way to run 2,000 pages a month. Per-page plans sit in the middle and trade a higher rate for a known number and someone else’s operations.
The math for 100,000 pages a month
Per token
We use one page profile for routine invoices — 2,200 input tokens and 400 output tokens — which covers a reasonably sized scan, a field list, and a JSON reply with header fields and line items. List prices at the time of writing, converted at $1 = €0.92: GPT-4.1 at $2 / $8 per million tokens, GPT-4.1 Mini at $0.40 / $1.60, Claude Sonnet 5 at $3 / $15.
Illustrative scenario. Routine invoice profile (2,200 input + 400 output tokens); list prices at the time of writing converted at $1 = €0.92; no caching, batch discount or volume tier. The 1.5× column adds retries and second passes on failed fields.
Model
Sticker per page
100,000 pages, sticker
100,000 pages at 1.5×
GPT-4.1 Mini
€0.0014
€140
€210
GPT-4.1
€0.0070
€700
€1,050
Claude Sonnet 5
€0.0116
€1,160
€1,740
Illustrative scenario. Routine invoice profile (2,200 input + 400 output tokens); list prices at the time of writing converted at $1 = €0.92; no caching, batch discount or volume tier. The 1.5× column adds retries and second passes on failed fields.
The 1.5× is not decoration. Failures cluster on the longest and messiest pages, and re-sending 10% of documents is never a 10% surcharge. If you have not measured your own retry rate, 1.3–2.0× is the range we plan with. The mix matters too: move 20% of the pile into the hard-page profile (8,000 + 800 tokens) and the GPT-4.1 month rises to about €1,460 and the Sonnet 5 month to about €2,390 at the same multiplier. High-resolution phone photos of receipts put pages into that profile without anyone deciding to.
Managed fixed price
Ækora’s published bulk document processing plans price the finished record. The Scale plan is €4,999 a month for 75,000 pages with overage at €0.09 per page. At 100,000 pages that is €4,999 + 25,000 × €0.09 = €7,249, or €0.0725 per page. That is a managed pipeline: a schema for each document type, validation rules (totals reconcile, VAT arithmetic checks, duplicate detection), an exception queue, and delivery as CSV, JSON or an API, with processing on local models so pages stay in the EU. It is not a token price and should not be compared to one directly.
Do-it-yourself reserved GPU
A compact document VLM does about 20 pages a minute on an A100-class card, so 100,000 pages is roughly 83 GPU-hours. Paid on demand at €1.80 an hour that is €150 of compute, or €0.0015 per page. Reserved for the month — a rented card or a small owned box, with power and a slice of ops — the line is about €600, which at 100,000 pages is €0.006 per page with the card busy around 12% of the time. Then add the people: someone has to write the schema, the validation, the retry logic, the exception UI and the ERP integration, and keep the box patched. If you assign half an engineer at a loaded €6,000 a month — replace that with your own figure — the DIY line is closer to €3,600 than €600.
Sensitivity: 25,000 to 500,000 pages a month
Illustrative scenario, routine profile, 1.5× retry multiplier on the API columns. Ækora column is published plan arithmetic: Pilot (€999 for 10,000 pages, €0.15 overage) at 25,000 pages; Scale (€4,999 for 75,000 pages, €0.09 overage) above that. At 250,000 and 500,000 pages Enterprise pricing is quoted on request; the figure shown is the Scale + overage arithmetic as an upper bound. DIY column is hardware only; add engineering time.
Pages / month
GPT-4.1 Mini
GPT-4.1
Claude Sonnet 5
Ækora managed
DIY reserved GPU
25,000
€53
€263
€435
€3,249
€600 + people
100,000
€210
€1,050
€1,740
€7,249
€600 + people
250,000
€525
€2,625
€4,350
€20,749 (Enterprise quoted)
€600 + people
500,000
€1,050
€5,250
€8,700
€43,249 (Enterprise quoted)
€600–1,200 + people
Illustrative scenario, routine profile, 1.5× retry multiplier on the API columns. Ækora column is published plan arithmetic: Pilot (€999 for 10,000 pages, €0.15 overage) at 25,000 pages; Scale (€4,999 for 75,000 pages, €0.09 overage) above that. At 250,000 and 500,000 pages Enterprise pricing is quoted on request; the figure shown is the Scale + overage arithmetic as an upper bound. DIY column is hardware only; add engineering time.
Three things stand out. First, per-token cost is linear, so at 500,000 pages the Sonnet 5 line is 20× the 25,000-page line while the reserved GPU barely moves; one card runs 500,000 routine pages in about 417 hours, under 60% of a month, though you would want a second card for headroom and reprocessing. Second, the managed column is the finished record, so the right comparison is not the token column but the token column plus the team. Third, the per-page rate of the managed pipeline falls with volume — €0.13 at 25,000 pages, €0.0725 at 100,000 — which is why Enterprise pricing exists above the Scale allowance.
Hidden costs on each side
On the per-token side
Schema and prompt engineering. Every document family needs a field list, examples, and a format the ERP accepts. A new supplier layout is a change request.
Validation code. Models produce confident wrong totals. Line items must sum, VAT must match the rate, dates must parse. None of that is in the token price.
Exception handling. If 5% of pages need a human, 100,000 pages is 5,000 reviews a month. Someone owns that queue and a tool to work it.
Integration and monitoring. Queues, rate limits, retries, dashboards, and the on-call when the vendor deprecates a model version.
Price and transfer risk. The rate card can change; the exchange rate moves; and sending supplier invoices to a US API is a data transfer you need a GDPR story for — see our guide to GDPR-compliant document AI.
On the fixed side
Idle hardware. A €600 card at 10,000 pages a month is €0.06 per page. Below the crossover, fixed cost is the expensive option.
Capacity ceilings and spikes. Reserved capacity wants a floor. A quarter-end backlog either queues, triggers overage, or needs a second card.
Ops time. Drivers, model updates, disk, backups, and the person who notices at 02:00 that the queue stopped. In a managed plan that is the vendor’s problem and priced in; in DIY it is yours.
Commitment. Monthly plans and rented cards are easy to leave; an owned box is a three-year decision that pays off only at high utilization.
The most common budget failure we see is none of the above. It is “one frontier model for everything”: routing every clean invoice to Sonnet or GPT-4.1 because it was the model the pilot used. Our article on why to stop paying OpenAI per invoice covers how that habit forms and how to split the traffic.
“Fixed cost per page” as a budgeting tool
Finance leads do not budget in tokens. They budget a line: pages expected × rate per page. A fixed cost per page turns a document pipeline into that line. At €0.0725 per page, a forecast of 100,000 pages is €7,249 and a forecast of 120,000 pages 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, and none of those are visible in a purchase order.
The same logic is why teams with in-house engineering still move to a reserved GPU once volume is steady: a €600 line is a €600 line. Our explainer on on-premise document AI and the 2026 guide to local OCR and document models describe what runs on that card. If the pages are not allowed to leave the EU — common in DACH finance and anything touching health or employee data — the fixed-cost route also settles the transfer question, and the EU AI Act’s high-risk obligations, in force since 2 August 2026, add a second reason to know exactly where processing happens. See what the EU AI Act means for document processing.
How to build the business case
Count real pages, not documents. Pull twelve months of invoice volume by page. Note the floor, the average and the quarter-end peak. Reserved capacity is sized on the floor; per-token is priced on the peak.
Profile a 200-page sample. Log input and output tokens per page on the API you use today. Split routine from hard. Do not average them into one number.
Measure the retry and human-touch rate. Run the sample once, re-send the failures, and count. That is your multiplier. If you cannot measure it, use 1.5× and say so.
Price the three shapes at your floor and your peak. Per-token at the measured multiplier; a per-page plan with its overage; a reserved GPU with a capacity check.
Add the people to every line. Build effort once, run effort monthly, exception review monthly. Use loaded costs from your own finance team.
Add the constraints. Residency, works-council or DPO requirements, AI Act classification, audit trail. A constraint can remove an option before the spreadsheet does.
Pick a rate to budget. Whatever the option, express it as cost per page at expected volume, with the assumptions written down next to it, so next year’s forecast is a multiplication and not a re-investigation.
Which shape wins at which volume
As a rule of thumb from the tables above, and assuming you count people on every line:
Under about 25,000 pages a month, or spiky volume: per-token, on the cheapest model that passes your validation, with a human on the exceptions. Do not buy hardware for a pile that arrives twice a year.
Steady volume, no ML team, and a budget owner who wants one number: a managed per-page pipeline. You pay more per page than raw inference and less than building and staffing the stack.
Steady volume above the crossover — roughly 50,000 pages against Sonnet 5, 85,000 against GPT-4.1 — with platform engineers on hand: a reserved GPU running local document models, with a frontier API kept for the exceptions.
If you are not sure which of your workflows sit above that line, that is the question our local AI consulting engagements answer: which daily processes have the volume and sensitivity to justify fixed capacity, and which should stay metered. The European buyer’s guide to AI document processing walks through the same choice from the procurement side.
Next step: export one month of API usage, or one month of scanner output, and price it at all three shapes with a 1.5× multiplier and your own people cost. Whichever line is cheapest, write the result down as a cost per page. That single number is what the budget conversation needs.