feld.ai vs. building it yourself — GPT or Claude for document AI?

We are not going to pretend otherwise. Published list prices, retrieved 20 August 2026, in USD per million tokens:

Both providers discount batch processing by 50%; cached input costs roughly a tenth of the standard input rate. A scanned A4 page sent as an image typically consumes low thousands of input tokens plus a few hundred output tokens for the structured JSON you want back. That puts a single page at approximately one cent on Claude Sonnet 5, and below that on the smaller models.

Scaled up: 100,000 pages a year is somewhere between $500 and $1,100 of raw inference. That number is real. It is also, in our experience, about 5% of what the system costs to run.

Four conditions. If all four hold, build it, and we mean that:

1. Low volume — hundreds of documents a month, not tens of thousands.
2. One stable document type from a small, known set of senders whose layouts do not churn.
3. No compliance burden — nothing an auditor will ask to reconstruct in three years.
4. Spare engineering capacity — a named person with time, not an intern's summer project.

Plenty of teams meet all four. For them a vendor is overhead. The trouble starts when a team meets three of four and assumes the fourth will sort itself out.

This is the actual answer to the original question. The model call is one step in a pipeline; the rest of the pipeline is the product.

None of these is impossible. All of them are months. The honest framing is that you are not choosing between $1,100 a year and a vendor invoice — you are choosing between two engineering roadmaps.

Sovereignty: you cannot buy EU-only inference from the frontier labs. The Anthropic API is global by default. The only geography control published on the rate card is US-only inference, which carries a 1.1x multiplier across all token categories. There is no first-party EU-only option. To pin inference to European regions you go through Amazon Bedrock or Google Cloud regional endpoints — which carry a 10% premium over global endpoints and place the workload on US-owned hyperscaler infrastructure. Read that as a procurement question rather than a technical one: EU data residency from the frontier labs is a paid add-on, delivered on infrastructure owned by US companies. For a great many workloads that is entirely fine. For supplier declarations, customs documents, contracts and personnel files, it is a conversation with your DPO that does not end quickly. Our GPUs are in Feldkirch. That is not a cheaper version of the same thing; it is a different thing.

Model churn: your cost model has an expiry date you do not control. 2026 supplied two clean examples. Claude Opus 4 and Sonnet 4 were retired on 15 June 2026, announced 14 April 2026 — anyone with a production pipeline pinned to them faced a forced migration and a complete re-validation cycle, unbudgeted and on someone else's timeline. And Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text. Per-token prices did not change. Effective cost per page did. A business case built on last year's numbers broke silently, without anyone changing a line of code. You are not buying a price; you are renting one from a vendor whose roadmap you cannot see. For a chatbot, fine. For a system your auditors depend on, every model change becomes a re-validation project.

Run the prototype. Seriously — it costs almost nothing and it teaches you what your documents actually look like, which is information you need either way. Then take the prototype and ask five questions of it:

1. What is its field-level accuracy across 200 documents, not 10?
2. When it is wrong, does it tell you it is unsure — and is that signal calibrated?
3. Who reviews the uncertain cases, in what interface, and how long does each take?
4. If an auditor asks in 2029 where a figure came from, what do you show them?
5. Who owns this when its author leaves?

If the answers come out well, you have saved yourself a vendor invoice. If they come out uncomfortable, you now know the scope — and we are happy to talk. For the legal grounding: EU-sovereign document AI. For our rates: pricing.

What does it cost to extract a page with GPT or Claude?

Very little. A scanned A4 page typically consumes low thousands of input tokens plus a few hundred output tokens. That puts a page at roughly one cent on Claude Sonnet 5 and below that on the smaller models. At 100,000 pages a year that is about $500 to $1,100 of raw inference. The number is real — in our experience it is also about 5% of what the system costs to run.

When is building it yourself the right decision?

When four conditions hold at once: low volume, one stable document type from a small known set of senders, no compliance burden an auditor will want to reconstruct in three years, and spare engineering capacity with a named owner. If all four hold, build it — a vendor is overhead.

What is not included in the API price?

OCR for scans and photos, layout and table fidelity, calibrated confidence scores, visual grounding, a human review interface, validation against master data and ERP, per-document-type prompt maintenance, an eval and regression suite, audit trail and retention, retry and failover logic, and accuracy-drift monitoring. None of it is impossible; all of it is months of engineering.

Can I get EU-only inference from OpenAI or Anthropic?

Not on their first-party APIs. The Anthropic API is global by default; the only published geography control is US-only inference, carrying a 1.1x multiplier. To pin inference to European regions you go through Amazon Bedrock or Google Cloud, where regional endpoints carry a 10% premium and place the workload on US-owned infrastructure. EU data residency from the frontier labs is a paid add-on delivered on someone else's infrastructure.

What happens when the model is deprecated?

Claude Opus 4 and Sonnet 4 were retired on 15 June 2026, announced 14 April 2026. Anyone with a production extraction pipeline pinned to them faced a forced migration and a full re-validation cycle. In addition, Claude 4.7 and later use a tokenizer producing roughly 30% more tokens for the same text: per-token prices did not change, effective cost per page did. A business case built on last year's numbers breaks silently.

feld.ai vs. ABBYY Vantage · feld.ai vs. Konfuzio · feld.ai vs. Nanonets · feld.ai vs. natif.ai · feld.ai vs. Rossum