A standard 50 pages document (~25,000 words / 33,325 tokens) costs $0.0117 to ingest and $0.012 to generate a comprehensive executive summary using NVIDIA Llama 3.1 Nemotron 70B.
Quick answer
A 50 pages document is modeled as 25,000 words or 33,325 tokens. With NVIDIA Llama 3.1 Nemotron 70B, ingestion costs $0.0117; a 500-token executive summary brings the modeled total to $0.012.
Method & trust
The page uses approximately 500 words per page, estimates tokens with the selected model tokenizer, and prices input plus a 500-token summary at the model's listed rates. OCR, tables, and formatting can increase the actual count.
Feed the entire 50 pages document into context to extract fields, entities, or answer queries.
Ingest 50 pages and generate a structured 500-token executive summary with key takeaways.
Translate or reformat the entire 50 pages document into an equivalent length output.
| Model | Provider | Ingestion (50 pages) | Cached Ingestion | Summary Cost | Context Limit |
|---|---|---|---|---|---|
| NVIDIA Llama 3.1 Nemotron 70B (Current) | nvidia | $0.0117 | $0.002916 | $0.012 | 128,000 |
| GPT-5.6 Terra | openai | $0.064 | $0.0064 | $0.07 | 1,050,000 |
| GPT-5.4 Workhorse | openai | $0.08 | $0.008 | $0.0875 | 256,000 |
| o3-mini | openai | $0.0352 | $0.0176 | $0.0374 | 200,000 |
| o4-mini | openai | $0.0352 | $0.0088 | $0.0374 | 256,000 |
| o1-mini | openai | $0.0352 | $0.0176 | $0.0374 | 128,000 |
Assuming a standard single-spaced document with ~500 words per page, a 50 pages document contains approximately 25,000 words, which translates to roughly 33,325 tokens using NVIDIA Llama 3.1 Nemotron 70B's tokenizer. Dense PDF tables, legal boilerplate, and OCR scans may increase this by 20–30%.
NVIDIA Llama 3.1 Nemotron 70B has a context window of 128,000 tokens. A 50 pages document consumes only 26.04% of its total window, easily fitting in a single prompt without requiring chunking or vector search.
Ingesting the document and outputting a concise 500-token executive summary costs $0.012. Using 24-hour async batch API queues drops this cost to $0.006007.
If your workflow repeatedly queries or chats with this same 50 pages document, prompt caching reduces subsequent turn input costs by up to 90% (to $0.002916 per turn instead of $0.0117).