LLM Cost by Use Case: OCR, Summarization & RAG

Document extraction (OCR) — from $0.340 per 1,000 pages

Assumptions: 1,250 input / 650 output tokens per page, 10% prompt-cache hit rate, batch API where available. Monthly column priced at 50,000 pages/month.

Which tier to pick: Budget models handle clean, single-column scans well. Complex layouts (tables, multi-column, handwriting) usually justify a mid-tier model — extraction errors are expensive to catch downstream. Verify document/vision input support per provider before committing; DeepSeek in particular should be tested on your own PDFs first.

Summarization — from $0.510 per 1,000 articles

Assumptions: 3,200 input / 300 output tokens per article, 5% prompt-cache hit rate, batch API where available. Monthly column priced at 100,000 articles/month.

Which tier to pick: Summarization is the most forgiving use case — budget models are usually sufficient for single-document summaries. Move up a tier for multi-document synthesis or when factual-consistency requirements are strict.

RAG generation — from $0.314 per 1,000 queries

Assumptions: 3,500 input / 400 output tokens per query, 60% prompt-cache hit rate, real-time (no batch). Monthly column priced at 100,000 queries/month.

Which tier to pick: The generator mostly rephrases retrieved evidence, so mid-tier models are the sweet spot: budget models hallucinate more on synthesis across chunks, frontier models rarely improve grounded answers enough to justify 3-10x the price. Spend the difference on better retrieval instead.

Generation cost only — embedding and vector-store costs are excluded (typically a small fraction at this scale; see methodology). Running extraction or RAG inside a multi-step agent? Price retries and chain depth with the agent cost calculator.

How much does LLM-based OCR / document extraction cost per 1,000 pages?

From $0.340 per 1,000 pages on DeepSeek DeepSeek V4 Flash to $21.94 on Anthropic Claude Fable 5 — a 65x spread for the same job (assumptions: ~1,250 input tokens per page image, 650 output tokens of structured text, batch API where available; prices verified 2026-07-14). A mid-tier pick like DeepSeek DeepSeek V4 Pro runs $1.06 per 1,000 pages and is the safer default for complex layouts.

What is the cheapest model for summarizing 1,000 articles?

DeepSeek DeepSeek V4 Flash at $0.510 per 1,000 articles (~3,200 input / 300 output tokens per article, batch API where available, verified 2026-07-14). Batch discounts reshuffle this ranking: models with a 50% batch discount can undercut nominally cheaper models without a batch API — check the full table for both orderings.

How much does RAG cost per month at 100,000 queries?

Generation cost ranges from $31.39/month on DeepSeek DeepSeek V4 Flash to about $96.46/month on a mid-tier model like DeepSeek DeepSeek V4 Pro (3,500 input / 400 output tokens per query, 60% prompt-cache hit rate on the system prompt and hot chunks). Embedding and vector-store costs are excluded — at this scale they are typically a small fraction of generation spend.

Why does the prompt-cache hit rate differ so much between use cases?

Because the share of repeated input differs. RAG re-sends the same system prompt and frequently-retrieved chunks on every query (we assume 60% cached). OCR and summarization feed mostly novel content — only the instruction template repeats (5-10%). Since cached input is billed at ~10% of list price (~1% on DeepSeek), cache-heavy use cases favor providers with strong caching more than the headline price table suggests.

Should I use a frontier model for OCR or summarization?

Rarely. Extraction and summarization are constrained tasks where mid-tier and budget models perform close to frontier models on standard content, at 5-70x lower cost. Frontier models earn their price on open-ended reasoning, complex synthesis and agentic work. Run a 100-document eval on a budget model first; upgrade only where it measurably fails.

Is classic OCR software cheaper than LLM-based extraction?

For plain text detection on clean scans, dedicated OCR services are usually cheaper and remain the right tool. LLM-based extraction wins when you need structured output in one step: key-value pairs, table extraction to JSON, cross-field validation, or handling messy layouts — replacing an OCR-plus-parsing pipeline with a single call. Compare against the total cost of that pipeline, not the OCR line item alone.