Document automation
PDF Summary Cost Planning
PDF size is only a rough signal. Text density, scanned pages, tables, images, and language all change how many tokens the model receives.
Why PDF estimates vary
A 20 MB PDF can be a text-heavy manual, a scanned contract, a presentation exported as PDF, or a report filled with images. These files have very different token counts after extraction. Text-heavy documents usually produce more tokens per megabyte than slide-like PDFs or image-heavy scans.
Language also matters. Korean, Japanese, and Chinese text can tokenize differently from English. The safest approach is to test a representative sample, then apply a buffer before estimating the full batch.
Good default assumptions
For text-heavy PDFs, estimate 30-45% of file bytes as useful text. For slide decks or scanned PDFs, useful text can be much lower unless OCR is applied. If OCR is required, include the OCR process and possible extraction errors in your project plan.
Output ratio
A short executive summary may be 10-15% of the input. A detailed chapter-by-chapter summary can reach 40-60% of the input. If the output includes tables, action items, risk analysis, and citations, use a higher output ratio in the calculator.
Example
If 100 MB of Korean PDF manuals become about 16 million input tokens after extraction and buffering, a 15% summary creates about 2.4 million output tokens. Running the job once is a batch cost. Re-running it every day is an operating cost.
Batch workflow advice
Summarize each PDF once, store the result, and reuse it in later workflows. Re-summarizing the same document for every user question is usually wasteful. For large libraries, summarize by section first, then produce a final digest from section summaries.