How much does it cost to summarize documents with AI?

The median model costs 0.83 cents to summarize a 20-page PDF. The cheapest measured model, GPT-6 Luna, comes in at 0.16 cents per summary, while the most expensive, Claude Fable 5.1, reaches 22 cents. The document is far larger than the summary, so input tokens drive most of the bill.

What a 20-page PDF costs

I ran the task on a 20-page PDF of 10,023 words (J. S. Mill's On Liberty). Across the 326 models I track, the median cost is 0.83 cents per task. The spread is wide:

• Cheapest: GPT-6 Luna at 0.16 cents per task ($1.57 per 1,000).
• Median: 0.83 cents per task.
• Most expensive: Claude Fable 5.1 at 22 cents per task ($222.89 per 1,000).

The table shows the five lowest-cost models with their token counts. Input tokens are the largest number in every row.

For the full ranking, see the PDF summary task page. To model your own volume, use the cost calculator.

#	Model	Cost per task		Text score	Tokens in / out	Basis
1	GLM 5.3 FlashZ.ai	0.10¢$0.99 / 1,000		1474	12,995 / 1,814	measured best value
2	DeepSeek V4.1 FlashDeepSeek	0.12¢$1.22 / 1,000		1475	12,970 / 1,992	measured best value
3	GPT-6 LunaOpenAI	0.16¢$1.57 / 1,000		1443	12,987 / 533	measured
4	Claude Haiku 5.5Anthropic	0.22¢$2.20 / 1,000		-	18,004 / 790	measured
5	Qwen3.8 FlashAlibaba Qwen	0.32¢$3.21 / 1,000		-	13,119 / 2,642	measured
6	GPT-5.4 NanoOpenAI	0.33¢$3.26 / 1,000		1401	12,987 / 533	estimated

Why input price matters most

A summary is short compared with the document it describes. In my data, visible output ranged from 148 to 857 tokens, with reasoning adding up to 2,141 tokens in the most reasoning-heavy runs. Input was about 13,000 tokens for most models and 18,004 for Anthropic models.

So the input price per million tokens sets the floor on cost. Two examples:

• GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 per million input tokens. Haiku still costs more per summary (0.22 cents against 0.16 cents) because it reads about 5,000 more tokens and writes more output.
• Claude Sonnet 5.5 lists at $2.00 per million input tokens. At 18,004 input tokens and 857 output tokens, the task costs 4.5 cents, far above the Luna figure.

When choosing a model for summarization, start with input price, then check output and reasoning behavior. Output matters more for tasks that generate long text, such as blog posts.

The tokenizer effect

Tokenizers do not count text the same way. For the same English document, the ratios (OpenAI = 1.00) are:

Anthropic's tokenizer produces about 39 percent more tokens than OpenAI's for the same text. That is why Claude Haiku 5.5 reads 18,004 tokens on this PDF while GPT-6 Luna reads 12,987. Both have the same input price, so the gap in input cost comes from the tokenizer.

A model with a lower list price can still cost more per document if its tokenizer splits the text into more pieces. Compare cost per task, not list price alone.

Cheapest options

These are the lowest-cost models in my results:

• GPT-6 Luna: 0.16 cents per task. Input $0.10 per million, output $0.50 per million. Measured.
• Claude Haiku 5.5: 0.22 cents per task. Same list price as Luna, but more tokens for the same text. Measured.
• GLM 5.3 Flash: 0.29 cents per task. Input $0.15 per million, output $0.50 per million. Measured.
• Qwen3.8 Flash: 0.32 cents per task. Input $0.15 per million, output $0.47 per million. Measured.
• Mistral Small 4: 0.33 cents per task. Input $0.15 per million, output $0.60 per million. Estimated.
• GPT-5.4 Nano: 0.33 cents per task. Input $0.20 per million, output $1.25 per million. Estimated.

Reasoning can change the picture. Qwen3.8 Flash and GLM 5.3 Flash used 2,141 and 1,270 reasoning tokens respectively, and these are billed as output. See the GPT-6 Luna page for an example model profile.

Expensive options

Frontier models cost more for the same task:

• Claude Sonnet 5.5: 4.5 cents per task ($44.58 per 1,000). Measured.
• Claude Opus 5.5: 8.9 cents per task ($89.16 per 1,000). Estimated.
• GPT-6 Astra: 16 cents per task ($155.27 per 1,000). Estimated.
• Claude Fable 5.1: 22 cents per task ($222.89 per 1,000). Estimated.

For many summarization workloads a cheaper model may be enough. Test quality on your own documents before paying the top-end price. You can also review the plans page.

How I measured

I ran each measured model once per task, with default settings, through OpenRouter, with output capped at 16,000 tokens. Other models use estimates, and the tables label which is which. Reasoning tokens are billed as output tokens, so they are included in the cost shown. Details are in the methodology.

Costs are in USD and reflect list prices as of 2026-10-11. Prices change, so re-check before budgeting.

FAQ

How much does it cost to summarize a 20-page PDF with AI?

Across the 326 models I track, the median is 0.83 cents per summary. The cheapest model, GPT-6 Luna, costs 0.16 cents per summary.

Why does the same document cost different amounts on different models?

Tokenizers differ, so the same text becomes different token counts. Anthropic's tokenizer produced about 39 percent more tokens than OpenAI's for the same English text. List prices and reasoning behavior also change the total.

What is the fastest way to lower summarization costs?

Choose a model with a low input price, since input tokens dominate this task. Then check how many reasoning tokens it uses, because those are billed as output.