What It Costs to Run an LLM on a 250-Page Case File: A 2026 Model Comparison

Model prices are published per million tokens, which says little to someone who has to budget for documents. A claims team, a law firm or an operations manager needs a different number: what one real file costs, and what a month of files costs. This comparison works that out for ten current models using the providers' own list prices and one realistic document.

For long business documents, the choice of model and the number of passes over the file change the model bill by a factor of a hundred or more, while OCR and human review often cost more than the model itself.

This article prices one 250-page case file across ten models from OpenAI, Anthropic, Google and Mistral, explains what moves the bill and compares the result with a real production system. It is not a quality benchmark, and it does not rank models on accuracy.

Key takeaways

• On October 26, 2026 list prices, one read of a 250-page case file costs from about $0.02 on the cheapest models to $1.75 on the most expensive flagship.
• Reading the file five times, once per topic, multiplies the cost almost five times, and prompt caching brings most of that back down.
• A mid-tier model such as Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon processes the file five times with caching for about $0.50, or roughly $500 per 1,000 files.
• OCR adds about $0.38 per file with Amazon Textract, and attorney or analyst review usually costs more per file than any model.
• A real five-assistant legal system built on earlier-generation models cost about $3–4 per case, close to what a top-tier model costs today without caching.
• Two scheduled price increases matter for 2027 budgets: Gemini 4 Argon doubles on December 1, 2026, and Gemini 3.8 Flash doubles on January 1, 2027.

The case file we priced

The example is a personal injury or insurance claim bundle of 250 pages: medical records, police reports, correspondence and handwritten notes. After OCR, we assume an average of 600 tokens per page, which gives about 150,000 tokens of text. Dense typed medical records sit above that average, and short handwritten notes sit below it.

Each request also carries about 5,000 tokens of instructions, examples and templates. We priced three ways of processing the file:

1. Single pass. The model reads the whole file once and writes a structured summary of about 4,000 tokens.
2. Five passes. Five specialized prompts each read the whole file and write one section of about 1,500 tokens, for 7,500 tokens of output in total. This mirrors systems that split the work by topic for accuracy.
3. Five passes with prompt caching. The same five passes, but the file is cached after the first request, so the following four reads are billed at the cached-input rate.

We also show a single pass through each provider's batch API, which costs half the standard rate in exchange for results within hours instead of seconds.

Cost per case file, by model

Prices are standard list prices in US dollars, taken from each provider's official pricing page on October 26, 2026.

The price per million tokens behind each row comes from the official pages for OpenAI, Anthropic, Google Gemini and Mistral. Input dominates every result, because a case file is long and the output is short.

How to read the table

Tiers matter more than providers

Models in the same tier cost about the same. Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all land at $0.35 for a single pass. The big differences are between tiers: a flagship model can cost from about 8 to almost 100 times more than a small model from the same provider for the same file.

Small models are almost free, but test them first

GPT-6 Luna, Claude Haiku 5.5 and Gemini 3.5 Flash-Lite process the whole file for a few cents. They work well for classification, splitting a bundle into document types and extracting simple fields. For synthesis across 250 pages of contradictory evidence, they need careful testing against your own cases before you rely on them.

Caching recovers most of the cost of multiple passes

Five passes without caching cost almost five times a single pass. With caching, five passes cost only about 1.4 to 1.7 times a single pass, because four of the five reads are billed at a fraction of the input price. If a design needs several passes for accuracy, caching makes that affordable.

What changes the bill

Number of passes and how much each pass reads

The biggest lever is how often the full file is sent to the model. A design that retrieves only the relevant pages for each topic, instead of sending all 250 pages every time, can cut input tokens by half or more. The trade-off is that retrieval can miss a relevant page, so it needs testing on real files.

Page images instead of OCR text

Many models can read PDF pages as images. That avoids a separate OCR step, but each page image typically uses more tokens than its extracted text, often 1,000 to 2,000 tokens per page instead of about 600. Sending images can double or triple the input cost, which is worth it mainly for handwriting, stamps and forms where layout matters.

Long-context price tiers

Some models charge more when a single request passes a size threshold. Claude Haiku 5.5 costs $0.10 per million input tokens for prompts up to 100,000 tokens and $0.50 above that, so a 155,000-token request pays the higher rate. Our table uses the higher rate for Haiku. Splitting the file into chunks below the threshold would cut that cost by up to five times, at the price of more requests.

Scheduled price changes

Google has announced that Gemini 4 Argon prices double on December 1, 2026, which would raise the five-pass cached cost from $0.48 to $0.97 per file. Gemini 3.8 Flash doubles on January 1, 2027, from $0.20 to $0.41 per file. Budgets for 2027 should use the new rates.

Data location surcharges

Keeping processing in a specific region can add to the bill. OpenAI lists a 10% surcharge for regional processing on newer models, and Anthropic charges 1.1 times the standard rate for US-only inference on its recent models. For legal, medical and financial files, these surcharges are often worth paying.

The full cost of processing one case

Model fees are only part of the cost of processing a file. For a 250-page bundle, the other parts usually look like this.

The conclusion is consistent across models. Once a capable model is chosen, the largest cost per file is the time a person spends checking the result. Spending a little more on a stronger model often pays off if it shortens that review.

Amazon Textract prices are for the first million pages a month in the US West (Oregon) region, and larger volumes cost less.

Reality check: what a real system cost

For a US personal injury law firm, .wrk built a tool that turns case bundles of about 250 pages into structured legal opinions using five specialized assistants, one per topic. Running on the models and prices available when it was built, the system cost about $3–4 per case bundle in model usage. That is close to what a top-tier model costs today for five uncached passes in our table, $3.25 for Claude Opus 5.5.

The comparison shows how prices have shifted. The cost of the strongest models has stayed in the same range, while mid-tier models now deliver work that once needed a flagship for a fraction of the price. A system built two years ago can often cut its model bill sharply by moving some steps to a newer mid-tier model and adding caching, after testing quality on real cases. Read the legal document processing case study.

Monthly budgets at different volumes

The table below combines model usage with OCR at Textract rates for three common designs. Review time and hosting are not included.

Gemini 3.8 Flash figures use the 2026 price, and its model cost doubles from January 2027. At these volumes, even the top-tier option costs less per month than a single full-time employee. The decision between models is therefore mostly about quality, review time and data handling, not about the model bill.

How to choose a model for long documents

1. Build a test set first. Collect 30–50 real files with outcomes approved by your experts, and score every candidate model against them.
2. Start with a mid-tier model. Mid-tier models usually give the best balance of quality and cost for synthesis across long documents.
3. Use small models for the simple steps. Splitting a bundle into document types, extracting dates and classifying pages rarely need a large model.
4. Use a top-tier model only where it changes the outcome. For example, the final synthesis of contradictory evidence, if testing shows a clear gain.
5. Turn on caching for multi-pass designs. It is the easiest saving available.
6. Use batch processing for work that can wait. Overnight processing of new files costs half the standard rate.
7. Re-test when prices or models change. Providers release new models and change prices several times a year.

For a broader view of what drives AI project budgets, see our guides to AI document processing and AI software for law firms.

Method

• Prices: standard-tier list prices in US dollars per million tokens, from the official pricing pages of OpenAI, Anthropic, Google and Mistral on October 26, 2026. Batch figures use each provider's published 50% batch discount.
• File: 250 pages, 600 tokens per page after OCR, 150,000 tokens in total.
• Instructions: 5,000 tokens per request.
• Output: 4,000 tokens for a single pass, 7,500 tokens in total for five passes.
• Caching: the file is cached once and read from cache four times. Anthropic cache writes are priced at 1.25 times the input rate. OpenAI, Google and Mistral caches are priced at their listed cached-input rates, and Google's small hourly storage fee is not included.
• Exclusions: DeepSeek publishes its prices in yuan rather than US dollars, so it is not included. Taxes, regional surcharges and volume discounts are not included.