PDF Summary Cost Calculator
How much does it cost to summarize a long PDF with AI? Enter page count and find out.
Summarizing 1 PDF of 300 pages each costs as little as U$ 0.0327 with Llama 4 Maverick.
| Model | Per PDF | Monthly Total |
|---|---|---|
| bestLlama 4 Maverick | U$ 0.0327 | U$ 0.0327 |
| GPT-5.6 Luna | U$ 0.0413 | U$ 0.0413 |
| Gemini 2.5 Flash | U$ 0.0721 | U$ 0.0721 |
| Nemotron 3 Ultra | U$ 0.0918 | U$ 0.0918 |
| Claude Haiku 4.5 | U$ 0.1922 | U$ 0.1922 |
| GLM-5.2 | U$ 0.2316 | U$ 0.2316 |
| Gemini 2.5 Pro | U$ 0.2944 | U$ 0.2944 |
| GPT-5.6 Terra | U$ 0.4133 | U$ 0.4133 |
| Claude Sonnet 5 | U$ 0.5767 | U$ 0.5767 |
| Kimi K3 | U$ 0.5767 | U$ 0.5767 |
| Claude Opus 4.8 | U$ 0.9612 | U$ 0.9612 |
| GPT-5.6 Sol | U$ 1.03 | U$ 1.03 |
Estimate based on 400 tokens/page and a summary output ~12% the length of the input. Scanned PDFs needing OCR will use more tokens.
How PDF summary cost is calculated
Summarizing a PDF with an LLM costs money on both ends: you pay for the input tokens the model reads (roughly the full text of the document) and for the output tokens it writes (the summary itself). This calculator estimates input size using an average of tokens per PDF page for dense text, then assumes the summary output is a percentage of that length — both assumptions are shown in the disclaimer below the result. Multiply that per-PDF cost by how many PDFs you plan to process to get a monthly estimate.
The math behind summarizing a long document
Summarization has an unusual cost profile: almost all the money goes to reading, not writing. A dense 300-page PDF runs in the neighborhood of 150,000 input tokens, while a thorough executive summary might be 3,000-8,000 output tokens — the input side outweighs the output side by a factor of twenty or more. That inverts the usual advice: for summarization workloads, the input rate is the number to shop for, and a model with cheap input and mediocre output pricing can beat a model with the opposite profile.
Real pipelines also carry overhead the naive estimate misses. Documents longer than the model's window get split into chunks with overlapping edges, and map-reduce style pipelines generate intermediate summaries that are themselves billed as fresh input on the next pass. Depending on chunk size and overlap, expect 10-30% more tokens than the raw document length suggests.
Single pass, chunking, or batch: three strategies
If the whole document fits in the context window, a single pass is the simplest and usually cheapest per document — no overlap overhead, no intermediate summaries. Its weakness is quality on very long inputs, where models tend to underweight material buried in the middle.
Chunked map-reduce works with any model regardless of window size and often produces more faithful section-level coverage, at the price of the token overhead described above. And when you have many documents rather than one big one, batch processing is the lever that matters: both major providers run asynchronous jobs at roughly half price, which for a nightly pipeline of thousands of PDFs is the difference between two invoices of very different sizes.
Cutting the cost of summarization at scale
Clean before you send: headers, footers, page numbers, boilerplate disclaimers and OCR artifacts are billed like content but add nothing to the summary. If users ask questions about a document rather than wanting a full summary, retrieval usually beats summarization — indexing the document and sending only relevant passages can cut input tokens by an order of magnitude.
Before committing to a pipeline, check whether your documents even fit in one request with the pages-per-context guide, and price the full run on the OpenAI cost calculator using your real page counts.
The same 100-page PDF, priced on every model
Summarization is input-heavy, so the input rate dominates and the model ranking looks different from a chat workload. Take a 100-page document at roughly 400 tokens per page: 40,000 input tokens, plus a 1,500-token summary out. At August 2026 list prices, one pass costs about $0.245 on GPT-5.6 Sol ($5.00 / $30.00 per million), $0.238 on Claude Opus 4.8 ($5.00 / $25.00), $0.098 on GPT-5.6 Terra ($2.00 / $12.00) and $0.143 on Claude Sonnet 5 ($3.00 / $15.00).
Further down: $0.048 on Claude Haiku 4.5 ($1.00 / $5.00), $0.066 on GLM-5.2 ($1.40 / $4.40), $0.024 on Gemini 2.5 Flash ($0.30 / $2.50), $0.0098 on GPT-5.6 Luna ($0.20 / $1.20) and $0.0089 on Llama 4 Maverick ($0.20 / $0.60). The 30 July repricing put Luna within a rounding error of Maverick on this workload — because input dominates, and the two now share the same $0.20 input rate.
One document is pocket change on any of them. The choice only starts to matter at volume: 10,000 documents a month costs about $2,450 on Sol, $98 on Luna and $89 on Maverick. Run the numbers for your own page counts above, and see the full cost breakdown for the strategy behind the model choice.
Frequently asked questions
Does this account for scanned or image-based PDFs?
Not directly. Scanned PDFs typically need OCR before an LLM can read them, which either adds a separate OCR cost or inflates the token count if you feed images directly to a vision-capable model. Treat the estimate here as a floor for scanned documents.
Why does page count matter more than word count?
PDFs vary a lot in density — a page of dense financial tables tokenizes very differently from a page of double-spaced prose. We use an average tokens-per-page figure as a practical approximation; for a specific document, run a sample through the token calculator above.
Can I reduce the cost of summarizing many PDFs?
Yes — batching requests where the provider offers an async discount, using a cheaper model tier for a first-pass summary, and caching repeated boilerplate across documents are the most common ways teams cut this cost at scale.
Is it cheaper to summarize in stages — a summary of summaries?
For documents beyond the context window it's often the only option, and it does cap per-request costs. But each stage loses detail, and intermediate summaries are billed as new input. If you need high fidelity, fewer, larger passes on a long-context model usually beat many small stages.
How do I estimate a PDF's token count before sending it?
Extract the text and paste a representative sample into the [token calculator](/en/token-calculator), then scale by page count. Dense text averages around 500 tokens per page, but slides can be a fifth of that and legal text can run well past it.
How accurate is 400 tokens per page?
It is a deliberate middle estimate for ordinary prose pages. A sparse page with headings and white space can land near 250 tokens, while a dense academic page with footnotes and tables can exceed 600. Scanned PDFs run through OCR are noisier still and often cost more. Use 400 for planning, then measure a representative sample of your own documents with the token calculator before committing to a volume budget.
Do images and charts in the PDF cost extra?
If you send them to a vision-capable model, yes — images are billed as additional input tokens, priced by resolution, and a single full-page figure can cost as much as several pages of text. If your pipeline extracts only the text layer, images cost nothing but their content is lost, which matters for documents where the charts carry the argument. Decide which one your use case needs before pricing it.
Does chunking a long document cost more than a single pass?
Usually a little more, because each chunk resends the instructions and each intermediate summary becomes input again in the final merge — overhead of roughly 10–25% depending on chunk size. In exchange you get better recall on long documents and you are not bound by the window limit. On a 200k-token document, chunking is not optional for a 200k-window model anyway.
Curious where these numbers come from? Read our pricing methodology.