How much does it cost to summarize a 100-page document with AI APIs?

Summarizing a 100-page document (roughly 75,000 tokens) costs approximately $0.006 with Gemini 1.5 Flash, $0.019 with Claude 3.5 Haiku, $0.023 with GPT-4o Mini, and $0.375 with GPT-4o. Batch processing 1,000 such documents monthly would cost $6–$375 depending on the model chosen, making Gemini Flash 60× cheaper than flagship models.

Can I use cheaper embedding models instead of LLMs for document summarization?

Embedding models alone cannot generate summaries—they only convert text to vectors for search or clustering. Summarization requires generative LLMs like Gemini Flash or Claude Haiku. However, you can use embeddings ($0.0001/M tokens) to first identify relevant sections, then summarize only those chunks with an LLM, reducing total input tokens by 70–90% and cutting costs proportionally.

Do AI APIs charge differently for prompt caching with long documents?

Yes. Claude's prompt caching reduces repeat input costs by 90% (cached tokens cost $0.025/M vs $0.25/M for Haiku) when processing multiple documents with the same instructions. Gemini offers context caching at $0.01875/M tokens for cached portions. GPT-4o has no native caching discount, making Claude and Gemini significantly cheaper for high-volume summarization with reused prompts.

What's the cost difference between streaming and batch API calls for summarization?

OpenAI's Batch API offers 50% discounts on input/output tokens compared to standard API calls—GPT-4o Mini drops from $0.30/M to $0.15/M for input. Gemini and Claude don't offer batch discounts but remain cheaper overall. For summarizing 10,000 documents monthly, OpenAI batching saves roughly $150–$300, though Gemini Flash non-batch still costs less at baseline rates.

USE CASE · VERIFIED 14 JUL 2026

What is the cheapest AI API for long-document summarization?

The cheapest AI API for long-document summarization is Google's Gemini 1.5 Flash at $0.075 per million input tokens (up to 128K context), followed by Claude 3.5 Haiku at $0.25 per million tokens. For documents exceeding 128K tokens, Gemini 1.5 Flash remains most economical at $0.30 per million tokens, making it 3–10× cheaper than GPT-4o or Claude 3.5 Sonnet for batch summarization workflows.

Summarization is the most input-skewed workload shape there is: you feed in a large document and get back a few hundred words. That asymmetry changes the cost math completely compared to a balanced chat workload — the input price, not the output price, is what actually determines your bill.

Worked example: 800M input, 5M output tokens/month

A realistic shape for summarizing long documents — reports, transcripts, contracts — at real volume.

GPT-4o mini$123/mo

Gemini 2.5 Flash$123/mo

Grok 4.1$162.50/mo

Gemini 3 Flash$415/mo

Claude Sonnet 5$1,650/mo

Gemini 3.1 Pro$1,660/mo

GPT-4o$2,050/mo

At this input-to-output ratio, the gap between the cheapest model shown and GPT-4o (the priciest one in this specific table) is nearly 17x — far wider than the roughly 6x gap you'd see on a balanced chat workload. This is the single biggest reason to actually calculate your specific token ratio instead of assuming a pricier model costs "a bit more." True flagship-tier models (Claude Opus 4.8, GPT-5.5, Claude Fable 5) would push that gap considerably further still — see the full price table for those figures.

Does summarization actually need a flagship model?

Often, no. Straightforward summarization — condensing a report into key points, extracting action items from a transcript — is exactly the kind of task budget-tier models handle well, because the skill required is compression and extraction, not deep multi-step reasoning. Reserve the mid or flagship tier for summarization that requires genuine judgment: legal document review where nuance matters, or synthesizing conflicting information across multiple sources into a coherent take.

Context window sets your ceiling, not just your cost

Gemini's 1M-token context window (both 2.5 Flash and 3 Flash) means a single call can hold an entire long report without chunking it into pieces — which matters for quality, not just convenience, since chunked summarization can lose connections between sections that only make sense read together. If your documents regularly exceed 100K tokens, check the context window before you check the price; a cheaper model that can't fit the document in one call may cost you more in workarounds than the price difference saves.

How we'd actually decide

Straightforward summarization at volume: Gemini 2.5 Flash — cheapest, and its 1M context handles most documents in a single call.
Summarization requiring real judgment (legal, financial, conflicting sources): Claude Sonnet 5 — worth the roughly 13x cost jump for tasks where getting it wrong is expensive.
Documents that regularly exceed 1M tokens: Gemini 3.1 Pro — the only tracked model with headroom to spare at 2M tokens.

Worked example uses standard (non-batch, non-cached) list pricing verified 14 July 2026. Use the calculator with your own document volume and length for an exact estimate.