Extraction estimates · prices checked 2026-11-03
What field extraction costs, by model
Estimated cost per 1,000 short documents (700 input and 120 output tokens plus estimated reasoning on the current model), against Levanto Sage 2’s measured $0.624 per 1,000 on our bench.
| Current model | $/M in / out | Est. per 1,000 docs | Keep ≥ for 30% saving |
|---|---|---|---|
| Claude Fable 5.1 | $10 / $50 | $53.00 | 31% |
| GPT-6 Astra | $10 / $50 | $28.00 | 32% |
| Claude Opus 5.5 | $4 / $20 | $11.20 | 36% |
| Claude Sonnet 5.5 | $2 / $10 | $10.60 | 36% |
| GPT-6.1 Sol | $2 / $10 | $5.60 | 41% |
| Gemini 4 Argon | $2 / $10 | $5.60 | 41% |
| GPT-6.1 Terra | $1 / $5 | $2.80 | 52% |
| Gemini 3.8 Flash | $0.75 / $3.75 | $2.10 | 60% |
Estimates, not bills. Levanto Sage 2’s token counts and accuracy (96.9% of field values on clean synthetic documents, 88.4% on messier ones) come from our bench run on 2026-11-03; the current model’s tokens are our assumptions. Discounts excluded.
Measure it on your own documents with the free extraction audit. For yes/no calls see classification cost by model.
Questions
Why are some models missing?
We only publish an extraction page where a hybrid with Levanto Sage 2 could save 30% with it keeping at most 70% of documents. GPT-5 mini is already about $1.02 per 1,000 documents, so a hybrid would need to keep nearly all of them.
Why does Levanto Sage 2 cost more here than for yes/no calls?
It returns every field with a probability, about 607 output tokens per document on our bench, billed at its $1 per M output price. Yes/no calls return almost nothing.