OCR API vs. Vision LLM: What 1,000 Pages Actually Cost

1. What OCR costs per 1,000 pages

Start with the service you would actually buy. Basic text recognition is a different purchase from invoice-field extraction. The familiar $1.50 figure is a paid page rate, not a promise that every document product from that vendor costs the same. These rows use the same records and default regions as our OCR calculator, reviewed on October 30, 2026.

If your entire monthly workload is 1,000 eligible Google pages, the page charge is zero. The next paid pages use the listed tier. Crossing a volume boundary does not reprice all earlier pages. For that reason, the rest of this article compares vision models with a $1.50 paid block, not with a fictional $1.50 Google bill for the free block.

AWS region differences and conditional new-customer trials also matter. We use the default paid region and leave conditional credits out. At high volume, use the calculator's graduated monthly total. A vision model that beats $1.50 might still lose to the cheaper OCR tier. Document parsing products can return more than text, so the table is a billing comparison, not a claim of identical features.

2. A vision model bills tokens, not pages

The input charge is what the model reads. The output charge is what it writes back. Both count. Sending a page image and asking for three fields has a different cost from asking for every word on that page, even though the upload is identical. A PDF is a container; it does not have one token count that works across providers.

For Claude and Luna, assume an A4 page at 150 dpi: 1240 x 1754 pixels. Claude's estimate combines an assumed 2,250 text tokens with a 28-pixel patch image count. Luna's estimate sends the page as an image at high detail and uses 32-pixel patches with the configured multiplier. It is not a native PDF estimate with a separately extracted text layer.

The Gemini values follow Google's documented PDF path at default media resolution. Google says text pulled from a native PDF is not billed on Gemini 3 models. That only holds for the PDF path, so don't assume it for image uploads. These are cost estimates from documented rules and stated page dimensions. We have not benchmarked recognition quality or measured tokens on a sample document collection.

Check the input rules against the provider's Gemini document guide, Claude PDF guide and OpenAI image guide. Your upload method, page dimensions and API usage report should decide the production estimate.

3. Output tokens can become the bigger bill

Here is the whole calculation. Prices are dollars per million tokens. For 1,000 pages, divide each per-page token count by 1,000 to express the total in millions. Use the same output length for every model so that the comparison answers a cost question rather than mixing four different tasks.

cost per 1,000 pages = (input tokens/page x input price
                      + output tokens/page x output price) / 1,000

700 output tokens/page x 1,000 pages = 0.7 million output tokens
1,500 output tokens/page x 1,000 pages = 1.5 million output tokens

The first scenario returns 700 tokens per page. That is an illustrative output budget, not a conversion from pages to words. Language, punctuation, tables and JSON keys change the count. We hold input constant and change only the answer length in the second chart. Neither scenario includes a separate instruction prompt.

Haiku and Luna beat the paid OCR baseline here. Both Gemini models cost more despite their smaller inputs. Output rates explain that reversal. There is one useful exception to the output-dominates rule: Haiku's input charge still exceeds its output charge at 700 tokens. Read both sides of the formula before deciding where to cut.

Now only Haiku stays below $1.50. The extra answer length is paid on every page. Asking for a transcript and then asking for a structured copy of that transcript writes much of the information twice. If the downstream system needs only fields, request those fields directly. Shorter output is useful only if it still contains everything the task needs.

4. The output length where each model hits $1.50

break-even output tokens/page
  = (1.50 x 1,000 - input tokens/page x input price) / output price

These rounded figures are an editorial snapshot for October 30, 2026. Below the line, the model costs less than the $1.50 paid block under our assumptions. Above it, it costs more. The figures are rounded to whole tokens, so don't read much into the last digit.

This is a budget boundary, not an accuracy ranking. A short reply with missing fields is not a saving. Nor does this table say Haiku can read every document well enough for your workflow. It tells you which output lengths are worth testing. Changing the input dimensions, the text-layer estimate or the OCR tier moves the boundary.

5. What the tables leave out

• Prompt tokens: instructions, schemas and examples add input beyond the page budget. Repeating a long schema on every request raises the bill.
• Thinking tokens: billed reasoning can add output-side usage beyond the visible answer. Check each model's settings and usage fields.
• Retries: a timeout, invalid JSON or failed extraction can lead to another paid attempt. Record cost per accepted document as well as cost per call.
• Accuracy: a cheap result that needs manual correction can cost more to use. Test real scans, small text and the fields that matter to you.

The basic OCR lanes compared here recognize text; they do not promise your business fields in your schema. OCR responses can also include geometry and other metadata, and parsing products differ. A vision LLM can return fields directly, which changes the job you are paying for. Our invoice OCR API cost guide separates text recognition from field extraction so those prices do not get mixed together.

Storage, upload handling, application code and human review are outside these API totals too. Count the pages actually submitted, including pages you did not need. Dropping a blank page before the call saves its processing cost without asking any model to produce a shorter or less complete answer.

6. Choose the output you need, then choose the API

For a large archive that needs searchable text, begin with basic OCR. Page billing is easy to forecast and the higher-volume tiers can reduce the paid rate. Check the monthly free allowance before assuming a small job has any page charge. Keep the original document available so a reviewer can check a doubtful transcription.

For fields, document questions or interpretation of layout, test a vision model. Give it the schema you need and measure actual output tokens. An estimate based on 700 tokens is not useful if your accepted answers average much longer. Compare successful extraction on the same documents before making a model choice from a price chart.

A hybrid can work too: run OCR first, then send only the recognized text to an LLM. That removes the image input from the second step but adds an OCR charge. It can also lose layout clues. Try it where plain text preserves the information; keep difficult pages on a separate path that still sees the image.

Take a small representative set, log input and output usage, and check the result against your acceptance rules. Then enter the observed output length in the OCR API Cost Calculator. Its vision model estimator lets you change the output budget while the OCR side applies page allowances and tiers. That is a better purchasing number than assuming every PDF page costs the same.

FAQ