Cheapest LLM for Document Summarization in 2026 — Real Cost Comparison

Document summarization is one of the cheapest LLM workloads. The output is short, the prompts are simple, and modern cheap models do it nearly as well as flagships. If your bill for summarization is over $100/month and you're not summarizing thousands of documents per day, you're using the wrong model.

What a Summarization Job Actually Costs

A typical document summary has this shape:

• Input: System prompt (200 tokens) + document text (varies)
• Output: Summary (typically 100-500 tokens)

The document text dominates. A 5-page document is roughly 2,500 tokens. A 20-page document is roughly 10,000. A 100-page report is roughly 50,000.

Cost Per Document by Size

Doc size	Input tokens	GPT-4o mini	GPT-4o	Claude Sonnet 4	Claude Opus 4
1 page (~500w)	700	$0.000165	$0.00275	$0.0033	$0.01650
5 pages (~2,500w)	3,200	$0.00075	$0.0125	$0.015	$0.075
20 pages (~10K w)	12,700	$0.00255	$0.0425	$0.051	$0.255
50 pages (~25K w)	31,700	$0.0063	$0.105	$0.126	$0.630
100 pages (~50K w)	63,200	$0.0126	$0.210	$0.252	$1.260
Book (~100K w)	125,700	$0.0252	$0.420	$0.504	$2.520

Even a 100-page report on GPT-4o mini costs about 1.3 cents. Summarizing 1,000 documents per day at this volume costs $13/day or about $390/month. The same workload on Claude Opus 4: $1,260/day or $37,800/month — 96x more.

How to Summarize Very Long Documents

Most cheap models have 128K-200K context windows. Gemini 2.5 Pro has 2M. For documents longer than your model's context window, you need a strategy:

Strategy 1 — Map-reduce. Split the document into chunks of 8K-15K tokens. Summarize each chunk independently. Concatenate the summaries. Run a final summarization pass over the combined summaries. Cost: roughly 1.3-1.5x the single-pass approach. Quality: usually slightly worse than single-pass, but works on any document length.

Strategy 2 — Long-context single pass. Use Gemini 2.5 Pro (2M tokens) or Claude Sonnet 4 (200K tokens). Pass the entire document in one call. Quality: best, because the model sees everything in one pass. Cost: high — a 200K input on Sonnet 4 is $0.60 per document.

Strategy 3 — Extract first, then summarize. Use a cheap model to extract just the important sections (intro, conclusions, headers, key paragraphs). Then summarize only those. Cost: lowest — typically 40-60% of single-pass. Quality: good if your extraction prompt is well-tuned.

Cost-Per-Document Across Strategies

For a 50-page (31,700 token) document on different models and strategies:

Strategy	GPT-4o mini	GPT-4o	Claude Sonnet 4
Single pass	$0.0063	$0.105	$0.126
Map-reduce (3 chunks)	$0.0089	$0.149	$0.179
Extract + summarize	$0.0034	$0.063	$0.076

Extract-and-summarize is usually the cheapest approach for long documents. Single-pass is cheapest if the document fits in context.

Quality Differences That Actually Show Up

The honest truth: for routine summarization, cheap models are nearly indistinguishable from premium models. Where differences appear:

• Long documents (50+ pages): Premium models maintain coherence better and forget less of the early content
• Technical documents: Claude Sonnet 4 and GPT-4.1 handle jargon and equations more accurately
• Multilingual content: Gemini 2.5 Pro tends to handle non-English content better than the others
• Structured output (sections, bullets, headings): GPT-4.1 and Claude Sonnet follow structure prompts more reliably

For news article summaries, blog post summaries, meeting notes, or short reports — the cheap tier is fine. Use the savings on a better RAG retriever or a faster vector DB.

The Recommended Stack for Summarization

1. Default model: GPT-4o mini for documents under 100 pages, Gemini 2.5 Flash for longer
2. Long documents: Gemini 2.5 Flash (1M context) or Gemini 2.5 Pro (2M context)
3. Output cap: 500 tokens unless you specifically need a long summary
4. Prompt structure: "Summarize the following document in N bullet points. Focus on [specific aspect]. Output format: markdown."
5. Quality check: if the summary is too short, too generic, or misses key points, escalate that prompt to GPT-4.1 or Claude Sonnet 4

Use the AI Cost Calculator with the "Document Summary" preset to model your real workload.