Stop paying OpenAI per invoice: a workflow-first method to reduce API costs
TL;DR
Per-token bills grow on four multipliers — volume, resolution, retries and model choice. The same routine invoice page costs €0.0014 on GPT-4.1 Mini, €0.0070 on GPT-4.1 and €0.0116 on Claude Sonnet 5 at list prices.
Routing beats model swaps. In an illustrative scenario of 20,000 invoices a month, a routed mix (Mini for routine pages, GPT-4.1 for hard ones) comes to about €120, against €281 all on GPT-4.1 and €461 all on Sonnet 5.
At 20,000 invoices a dedicated €600 GPU does not pay on the invoice line alone. Local wins once the same GPU also carries receipts, triage and search, or once EU residency rules the API out.
Keep OpenAI or Anthropic for four things: the exception queue, open-ended low-volume work, high-stakes reasoning on redacted text, and overflow — always with personal data stripped and a DPA in place.
A 30-day plan — export, inventory, measure, classify, route, pilot, build the exception path, decide — is enough to move the first workflow off the meter.
Questions people ask
How do I reduce OpenAI API costs for document processing?
Route by workflow instead of swapping models. Inventory every recurring document task, measure tokens per unit and monthly units from your usage export, classify each by sensitivity and variability, then place it on the cheapest tier that meets its bar: a local specialist model for high-volume stable work, a mini API for forgiving tasks, a frontier API for rare open-ended work, and a person below the confidence threshold.
Is a local model cheaper than OpenAI for 20,000 invoices a month?
Not on the invoice line alone. In an illustrative scenario at list prices, 20,000 invoices (about 25,000 pages) cost roughly €281 a month all on GPT-4.1 and €120 on a routed Mini/GPT-4.1 mix, while a dedicated €600 GPU comes to about €626. Local becomes the cheapest line when the GPU is shared across several workflows, or when the invoices cannot leave the EU.
What should stay on OpenAI or Anthropic after moving to local models?
Four categories: the exception queue from the local model (a new layout, an unsampled language, a handwritten note), open-ended low-volume work such as proposal drafts and one-off analyses, high-stakes reasoning on pseudonymized text, and overflow during volume spikes. In every case strip personal data first, use an EU data-residency option where the provider offers one, and keep a GDPR Art. 28 processing agreement in place.
How do I measure what a workflow costs on the API?
Export a month of usage by API key or project, and map each key to a workflow. Then log input and output tokens for 200 real units of that workflow, splitting routine units from hard ones instead of averaging them, and record the retry and human-touch rates. Multiply by a steady monthly volume rather than a backlog. That number, not the rate card, is what you compare against a local GPU line.
Why do per-token bills grow faster than document volume?
Because tokens, not documents, are what is invoiced, and four multipliers compound: volume across every workflow on the key, image resolution (a phone photo can cost 3.6× the input tokens of a routine scan), retries and second passes at 1.3–2.0×, and model choice, where a frontier model costs about eight times a mini model per page. A high-resolution page on a frontier model with a retry can cost 35× a clean page on a mini model.
What does a local AI consulting engagement deliver?
Three things. A map of the week: every recurring workflow with owner, volume and current tool. A score for each on volume, sensitivity, accuracy bar and whether a frontier API is still worth the premium, plus a routing map and a cost line per route. And a pilot of the first local model on a real workload — usually invoices or triage — with measured accuracy, throughput and human-touch rate to support the hardware decision.
Want this worked out on your documents?
We will price a real sample against OpenAI or Anthropic and tell you whether Local AI consulting or a local model on your hardware is the cheaper first step.
The fastest way to reduce OpenAI API costs is not a cheaper model or a shorter prompt. It is routing. Stop sending every recurring document and text task to a frontier model, and put each workflow on the cheapest tier that meets its quality and residency bar: a local specialist model, a mini API, a frontier API, or a person. In a typical operations department, a handful of high-volume, repetitive workflows — invoices, receipts, inbox triage, internal search — generate most of the tokens, and none of them need GPT-4.1 or Claude Sonnet 5 on every unit. Move those, keep the frontier API for the exceptions, and the bill stops tracking your volume.
This is the method we use in local AI consulting engagements, written out in full. It is not proprietary. You can run it yourself in 30 days with a usage export and a 200-page sample.
How a per-token bill grows
OpenAI and Anthropic do not invoice per invoice. They invoice per token, and four multipliers sit between “we process 20,000 invoices” and the number on the card statement:
Volume. Units per month, across every workflow that shares the API key. Invoices are rarely alone; receipts, supplier emails and the odd contract ride on the same account.
Resolution. A routine invoice page costs about 2,200 input and 400 output tokens. A high-resolution scan or a phone photo tiled for a vision model lands in the hard-page profile at 8,000 + 800 tokens — 3.6× the input before the model reads a single field.
Retries. Failures cluster on the largest, messiest pages. Re-sending the misses and running a second validation pass typically adds 1.3–2.0× to the sticker.
Model choice. At list prices at the time of writing, a routine page costs €0.0014 on GPT-4.1 Mini, €0.0070 on GPT-4.1 and €0.0116 on Claude Sonnet 5. That is an 8× spread for the same page.
The multipliers compound. A phone-photographed invoice sent to Sonnet 5 with a second pass costs about €0.0331 × 1.5 ≈ €0.05 — roughly 35× the Mini sticker for a clean routine page. Nobody chose that number. It is what “use the best model for everything” produces once a scanner is replaced by a mobile app.
The workflow-first method
Cost optimization at the model level — swap GPT-4.1 for Mini, cache the system prompt — saves a percentage. Cost optimization at the workflow level changes the shape of the bill, because it removes whole categories of work from the meter. Four steps.
1. Inventory the recurring work
List every document or text task that repeats weekly: supplier invoices, expense receipts, support and AP inboxes, contract intake, HR forms, internal search, translations, first drafts. Write down the owner, the current tool, and the API key it runs on. Most teams find eight to fifteen workflows, and two or three they had forgotten were on the API at all.
2. Measure tokens per unit and units per month
Export a month of usage from the OpenAI or Anthropic dashboard, split by key or project. For each workflow, log input and output tokens for 200 real units, and keep routine and hard units apart rather than averaging them into one comforting figure. Record the retry rate. Then take a steady monthly volume, not last quarter’s backlog. Hardware wants a floor; an API wants a spike.
3. Classify by sensitivity and variability
Sensitivity: does the unit contain personal data (GDPR Art. 6), special-category data such as health information (Art. 9), supplier bank details, or contract terms you are obliged to keep in-house? Must it stay in the EU? Variability: is the task a fixed schema on a stable set of layouts, or open-ended reasoning on something new each time? A local specialist model is very good at the first and weak at the second. A frontier API is the reverse, at eight to ten times the price per page.
4. Route each workflow to one of four tiers
Local specialist. High volume, stable pattern, any sensitivity. An OCR or document vision model, or a small text model for classification, on a GPU you control.
Mini API. Moderate volume, low sensitivity, forgiving accuracy bar. Cheap — but measure the miss rate on your own pages before you trust it.
Frontier API. Low volume, high variability, low sensitivity. Also the exception queue from the local tier, with personal data removed first.
Human. Anything below the confidence threshold where an error costs more than the two minutes a person needs to check it.
Typical workflows and where they belong
Default routing for common recurring workflows. Sensitivity follows GDPR categories; volume and variability are typical for a mid-sized European operations or finance team and should be replaced by your own inventory.
Workflow
Volume pattern
Sensitivity
Variability
Recommended route
Supplier invoice extraction
High, steady
Medium (bank details, contacts)
Low
Local specialist; frontier API or human on exceptions
Receipts and expense claims
High, month-end spikes
Medium (employee data)
Low–medium
Local specialist
Prescriptions, claims, medical forms
Medium–high
Very high (Art. 9)
Low–medium
Local only; human on exceptions, no API
AP and support inbox triage
High, continuous
Medium
Medium
Local small LLM classifier; human below threshold
Contract intake and clause review
Low–medium
High
High
Local long-context model; frontier API only after redaction
Internal knowledge search and Q&A
Medium, continuous
High (internal documents)
Medium
Local embeddings plus local LLM
Routine document translation
Medium
Medium
Low
Local small model, or mini API if volume is low
Sales and proposal drafting
Low, ad hoc
Low
High
Frontier API
Ad-hoc research and analysis
Low, irregular
Low
Very high
Frontier API
Default routing for common recurring workflows. Sensitivity follows GDPR categories; volume and variability are typical for a mid-sized European operations or finance team and should be replaced by your own inventory.
The table is a starting map, not a verdict. A contract workflow with 50 documents a month scores differently from one with 5,000; a support inbox in one language differs from one in nine. The guides to on-premise document AI and the current local OCR and document models cover what the local tier can actually do, so you route on evidence rather than hope.
What 20,000 invoices a month cost on each route
Take a finance team processing 20,000 supplier invoices a month. Most are one page; some run to two or three, so call it 25,000 pages: 22,000 routine and 3,000 hard (stamped, multi-column, or photographed). Five routes, same pages.
Illustrative scenario, assumptions stated here: 20,000 invoices = 25,000 pages a month (22,000 routine at 2,200 + 400 tokens, 3,000 hard at 8,000 + 800). API rows use list prices at the time of writing, $1 = €0.92, and a 1.3× retry multiplier. Local rows use a €600/month reserved GPU line and send 5% of pages (1,250) to GPT-4.1 at the hard-page rate. No engineering time is included in any row.
Scenario
Routing
Monthly cost
What drives it
A. All on Claude Sonnet 5
Every page to Sonnet 5
≈ €461
€355 sticker × 1.3 retries
B. All on GPT-4.1
Every page to GPT-4.1
≈ €281
€216 sticker × 1.3 retries
C. Routed mix
Routine to GPT-4.1 Mini, hard to GPT-4.1
≈ €120
(€31 + €62) × 1.3; depends on Mini’s miss rate on your pages
D. Local, dedicated GPU
All pages local; 5% exceptions to GPT-4.1
≈ €626
€600 reserved GPU + €26 API
E. Local, GPU shared by four workflows
As D, invoices carry a quarter of the GPU
≈ €176
€150 share + €26 API
Illustrative scenario, assumptions stated here: 20,000 invoices = 25,000 pages a month (22,000 routine at 2,200 + 400 tokens, 3,000 hard at 8,000 + 800). API rows use list prices at the time of writing, $1 = €0.92, and a 1.3× retry multiplier. Local rows use a €600/month reserved GPU line and send 5% of pages (1,250) to GPT-4.1 at the hard-page rate. No engineering time is included in any row.
Three things to read off the table. First, at this volume the invoice line alone does not justify a dedicated GPU: route D is the most expensive row. A reserved €600 GPU only overtakes GPT-4.1 on routine pages at about 85,000 pages a month, and Claude Sonnet 5 at about 50,000. Second, the cheapest API answer is routing, not a model swap: the routed mix in row C is less than half of row B and a quarter of row A. Third, local becomes the cheapest line the moment the GPU is shared. Row E is the realistic case — the same card that reads invoices in the morning runs receipts, inbox triage and internal search for the rest of the day, and the €600 is spread over the week rather than charged to one workflow.
Row C also carries a risk the others do not. If GPT-4.1 Mini drops line items or misreads VAT on your supplier mix, the “failed” pages go to GPT-4.1 anyway and the row drifts toward B. Measure the miss rate on 200 of your own invoices before you rely on it. And if the invoices contain data that cannot leave the EU, rows A to C need a transfer basis under GDPR Art. 44–49 before they are options at all — see GDPR-compliant document AI for what that entails.
For teams that want the local line without running the GPU, our bulk document processing plans price the same work as a managed pipeline — schema, validation and delivery to CSV, JSON or an API — starting at €999 a month for 10,000 pages. That is a different product from raw inference, and the comparison worth making is against your own engineering time, not against €0.0015 a page.
What to keep on OpenAI or Anthropic
“Stop paying per invoice” does not mean “cancel the account.” Frontier APIs earn their price on a specific set of work:
The exception queue. The 5% of pages a local model flags as low confidence — a new supplier layout, a language you have not sampled, a handwritten note across the total. Send the page, not the pile.
Open-ended, low-volume work. Proposal drafts, one-off analyses, an unusual customer letter. A GPU bought for 40 tasks a month idles for 29 days.
High-stakes reasoning on redacted text. A clause that has to be right, after names, amounts and identifiers have been pseudonymized.
Overflow. A quarter-end spike beyond what the reserved GPU can absorb. The API is elastic; hardware is not.
For all four, the same rules apply: strip personal data before the call, use an EU data-residency option where the provider offers one, and have a data processing agreement (GDPR Art. 28) in place. Special-category data under Art. 9 — prescriptions, claims, medical forms — does not go to the API at all. It goes to a person.
A 30-day plan to move the first workflow
Days 1–3: export and attribute. Pull a month of usage from OpenAI and Anthropic by key or project, and map every key to a workflow and an owner. Unattributed spend is usually the first surprise.
Days 3–5: inventory. List the recurring workflows with monthly units, the model they run on, and the accuracy bar the business actually needs.
Days 6–10: measure. For each workflow, log tokens per unit on 200 real units, split routine from hard, and record the retry and human-touch rates.
Days 8–10: classify. Tag each workflow for personal data, Art. 9 data, contractual restrictions and EU residency. This step is what makes the routing defensible to a DPO or a works council.
Days 11–12: route. Assign each workflow to local, mini API, frontier API or human, and write the routing map down. Rank candidates for the local tier by monthly tokens × sensitivity.
Days 13–20: pilot. Run the top workflow on a local model — usually invoices or triage — against the same 200 labeled units. Measure field accuracy, throughput and the share of units that fall below the confidence threshold.
Days 21–25: build the exception path. Set the threshold, send low-confidence units to a frontier API with personal data removed or to a review queue, and log what leaves the network.
Days 26–30: decide. Compare the pilot’s cost per unit with the export. Choose between a dedicated GPU, a shared GPU across the week, or a managed pipeline, and schedule the next two workflows.
How a consulting engagement runs this
Our engagement is the plan above with the measuring done by people who have done it before. Three phases. Map the week: we sit with operations, finance and support and list the work that repeats — documents, inboxes, search, drafts, reviews — with volumes and owners. Score local versus cloud: each workflow gets a score on volume, sensitivity, accuracy bar and whether OpenAI or Anthropic is still worth the premium, and you get the routing map. Pilot the first win: we stand up one local model on a real workload, measure cost, quality and the handoff, and hand over the numbers you will need for the hardware decision.
The output is deliberately boring: a table of workflows, a route for each, a cost line per route, and one pilot with measured accuracy. If the local tier is the right home for your documents, the next step is either local document models deployed on your hardware or bulk processing on ours. If the routed mix in row C is the right answer for your volume, we will say so — it is a cheaper engagement and a smaller change. The European buyer’s guide to document AI lays out the choices on that side.
What to do this week: export last month’s token usage, split it by workflow, and find the two workflows that account for most of it. Price those two on the routed mix and on a shared €600 GPU line. Whichever number is lower, that is the change to make first — and the invoice bill will stop growing with your invoice count.