LLM PDF Processing Cost: What 89,000 Documents Really Cost

Last updated: October 2026

Key takeaways

• A builder on Indie Hackers reported a total API bill of $0.77 for a project that read 89,000 planning decisions.
• The early version used a small hosted model. Newer versions run local models on a 2017 PC, which shifts cost from tokens to time and hardware.
• Per-token price is only one line of your document pipeline. Retries, review and re-runs often cost more than the tokens themselves.
• For SaaS founders, the number that matters is cost per finished document, because that is what you can price against.

How much does it cost to process PDFs with an LLM?

The short answer: far less than most founders fear, as long as the task is extraction and classification rather than long creative output. Reading is cheap because input tokens are priced low and the output is usually a short structured record.

A recent Indie Hackers post is a good real-world data point. The author reports a total API bill of $0.77 for the whole project, which reads planning notices from council websites. The first version used Claude Haiku through the API. The current version runs local models through Ollama (the post names Qwen3.5 9B) on a 2017 PC with a GTX 1060, and the author counts well over 34,000 local model calls.

Two details from the post matter for anyone building a paid product. First, a full re-read of the dataset takes roughly 20 hours on that hardware. Second, the author compares each new version on a fixed test set of 480 notices before promoting it. Cheap tokens did not remove the need for evaluation work.

Why is the API bill so small and is it the whole cost?

It is small because document reading is an input-heavy, output-light workload. You send a lot of text in and get back a few hundred tokens of structured data. Input tokens on small models are among the cheapest things you can buy in AI.

But the $0.77 is not the whole cost, and the post is clear about why. The newer pipeline moved the work off the API entirely. That means the dollar bill dropped while other costs appeared: a machine that has to be owned and powered, a re-read that takes most of a day, and the time spent comparing versions. None of those show up on an invoice, and all of them show up in your margin if you sell this as a service.

What does it cost to process 50,000 documents a month?

Here is an illustrative calculation. These rates are assumptions for a small, cheap hosted model, not a quote from any provider, so check the current numbers on the Anthropic pricing page or your own provider before you plan around them.

Say you process 50,000 contracts a month. Each is 10 pages at about 700 tokens per page, so 7,000 input tokens per document. The model returns about 300 tokens of structured fields.

| Line item | Volume per month | Illustrative rate | Monthly cost | |---|---|---|---| | Input tokens | 350M | $0.25 per 1M | $87.50 | | Output tokens | 15M | $1.25 per 1M | $18.75 | | Total tokens | | | $106.25 | | Cost per document | | | about $0.0021 |

If you charged a customer $0.05 per document, the token cost alone would leave a gross margin near 96 percent. That looks wonderful, and it is exactly why so many founders stop the analysis here.

What costs does the token bill leave out?

The token line is the easy part. The costs that decide whether your document product makes money sit around it.

| Cost layer | What it looks like | Who feels it | |---|---|---| | Retries | Malformed JSON, truncated pages, timeouts | Your token bill, multiplied | | Human review | Spot checks on low-confidence extractions | Your support or ops payroll | | Re-runs | Reprocessing everything when the prompt or model changes | Your compute or API bill, again | | Hardware or rental | GPU you own or rent for local models | Fixed cost, regardless of volume | | Evaluation | A fixed test set and a rule for promoting versions | Engineering time |

A useful habit is to track cost per finished document: total spend on tokens, retries, review and compute, divided by documents that passed your quality bar. If a 20 percent retry rate adds 20 percent to the token line, you barely notice it at $0.0021. If review touches 5 percent of documents at a few dollars of human time each, review becomes the biggest line. For a deeper look at this shift in thinking, see our breakdown of cost per token versus cost per task.

Should you run local models or use an API for document processing?

It depends on volume, speed and how sensitive the data is. The Indie Hackers builder had a hobby-scale workload, free hardware and no deadline, so local made sense. A SaaS with customers waiting for results is in a different position.

An API gives you speed, no hardware to manage and a cost that scales smoothly with usage. Local models give you a flat cost that gets cheaper per document as volume rises, but you take on reliability, throughput limits and maintenance. The crossover point depends on your utilization, and we walk through the math in self-host versus API LLM cost breakeven. If your documents feed a retrieval workflow, the RAG app cost calculator page shows how ingestion cost fits next to query cost.

How should you price a document-processing feature?

Price against cost per finished document with a healthy buffer, not against the raw token rate. Because token cost per document can be a fraction of a cent, founders are tempted to give the feature away. That is fine as a growth lever, but only if you have sized the retry, review and re-run layers first. Otherwise a cheap feature quietly becomes a support burden.

A simple rule: take your measured cost per finished document, add a buffer for model price changes and heavy users, then decide whether to charge per document, per page or through included credits.