LLM OCR pricing: a calculator for your real cost per document

Vision LLMs are priced in tokens, but you don't think in tokens — you think in documents and monthly bills. That gap is why teams routinely under- or over-estimate the cost of LLM-based extraction by an order of magnitude, then get surprised by the invoice.

The calculator below closes the gap. Put in your volume, a rough token estimate, and the per-million-token prices from your provider's pricing page, and it gives you the cost per document, per 1,000 documents, per month, and per year. The arithmetic is shown in full underneath so you can check every number yourself.

The calculator

The token and price defaults above are placeholders, not quotes. Token counts vary by model, image resolution, and prompt; prices change. Always pull the current per-million-token rates from your provider's pricing page and a token count from a real document before trusting the total.

The formula, in full

There's no magic here — just unit conversion. The cost of running one document through a vision LLM is the input cost plus the output cost:

input_tokens_per_doc  = pages_per_doc × input_tokens_per_page
output_tokens_per_doc = output_tokens_per_doc          (the JSON you get back)

cost_per_doc = (input_tokens_per_doc  ÷ 1,000,000) × price_per_1M_input
             + (output_tokens_per_doc ÷ 1,000,000) × price_per_1M_output

monthly_cost = cost_per_doc × documents_per_month
annual_cost  = monthly_cost × 12

Token prices are almost always quoted per million tokens, which is why every token term is divided by 1,000,000. That single division is where most back-of-the-envelope estimates go wrong — off by a factor of a thousand in either direction.

A worked example

Using the calculator's default placeholders — so you can check the math against the widget above:

Step by step:

• Input cost per document = 1 × 1,200 ÷ 1,000,000 × $2.50 = $0.0030
• Output cost per document = 350 ÷ 1,000,000 × $10.00 = $0.0035
• Cost per document = $0.0030 + $0.0035 = $0.0065
• Per 1,000 documents = $0.0065 × 1,000 = $6.50
• Per month = $0.0065 × 10,000 = $65.00
• Per year = $65.00 × 12 = $780.00

So under these assumptions, extraction costs $65/month. Now bump the output price to $30/1M (a reasonable difference between a budget and a premium model) and the per-document output cost triples to $0.0105, pushing the monthly total to $135. The model you pick — and the verbosity of the JSON you ask for — moves the bill far more than your document count does.

What the calculator deliberately leaves out

To keep the number honest and not falsely precise, a few real-world factors sit outside the formula. Account for them when the stakes are high:

• Retries and failures. Timeouts, rate limits, and malformed JSON mean some documents get sent more than once. A 5–10% retry rate is common; multiply accordingly.
• Prompt caching and batch discounts. Some providers discount cached input tokens or batched (asynchronous) requests substantially. If you qualify, your real input cost can be well below the list price.
• Image tiling. High-resolution pages are often split into tiles, each billed as input tokens — so a "1-page" document can cost several pages' worth of input. Measure tokens on a real page, not a thumbnail.
• Multi-pass pipelines. Classify-then-extract, or extract-then-verify, means more than one model call per document. Add a row of cost for each pass.

Cost is one axis of three

The cheapest model that gets the answer wrong is not cheap — it's a liability with a discount. Cost only means something next to accuracy and latency, which is why you should measure all three on the same documents before committing. See how to choose an OCR and extraction model for the full framework, and how OCR Bench scores extraction for how the accuracy half is graded.

Want the real per-document cost and accuracy for your documents, side by side? Start free, run a benchmark, and read all three numbers off one scorecard.