What counts as schema and instruction overhead?

Anything sent with every document — the JSON schema or field list you want filled, formatting instructions, and any few-shot examples. Because it repeats on every call, it can dominate cost when documents are short.

Why include a retry rate?

Real extraction pipelines re-run a fraction of documents when output fails validation, returns low confidence, or violates the schema. An 8% retry rate means effective calls per document are 1.08x, which the estimator factors into both per-document and total cost.

Which model should I pick for extraction?

Cheaper, faster models like GPT-4o mini, Haiku or Gemini Flash are often accurate enough for well-defined field extraction, and at scale the savings are large. Reserve premium models for ambiguous or high-stakes documents.

How do I cut extraction cost at scale?

Trim the schema and instructions to the minimum, batch documents where the API supports it, use a cheaper model with validation, and only escalate failed documents to a premium model — rather than running everything on the expensive one.

Is anything uploaded?

No. The estimate is computed entirely in your browser. No document content or figures are sent anywhere.

Data Extraction Pipeline Cost Estimator

Budget a document extraction pipeline before you run it

Extracting structured fields from thousands of invoices, contracts or records with an LLM can be cheap or eye-watering depending on document size, schema overhead and how often you retry. This estimator gives you a defensible per-document and total cost so you can size a pipeline — or compare models — before processing a single file.

How it works

Each document costs ((doc_tokens + schema_tokens) × input_price) + (output_tokens × output_price). The schema and instruction tokens are added to every document because you resend them on each call, which is why short documents with a big schema can cost more than you expect. The estimator then applies your retry rate: an 8% retry rate inflates effective calls per document to 1.08×, capturing the cost of re-running failed or low-confidence extractions. Multiply by your document count and you have the total pipeline cost.

Tips to keep extraction cheap at scale

Shrink the schema. Every redundant field description and example is billed on every document. Keep instructions tight.
Use a cheaper model with validation. A mini/flash model plus a schema validator often beats a premium model on cost per correct extraction.
Escalate, don’t blanket. Run everything on the cheap model and only retry the failures on a stronger one, rather than paying premium prices everywhere.
Batch where you can. Batch APIs and longer prompts that pack multiple records can cut per-document overhead — just watch context limits and accuracy.