What generating synthetic training rows with an LLM costs.
Bulk generation is cheap; the tokens lost to rejected and duplicate rows are the real spend. A tight prompt and a cheap generator win. Size the calls on the LLM token cost calculator and compare against fine-tuning.
The Synthetic Data Generation Cost Calculator estimates what it costs to produce an artificial dataset with a language model. You enter the number of rows to generate, the tokens per row (prompt plus output), and the price per 1 million tokens, and it multiplies these to project total spend. It also factors in a reject/retry percentage, since generated rows that fail validation or quality checks must be regenerated, adding tokens you actually pay for. The main cost drivers are therefore row count, token weight per row, model price, and how often output is retried.
The key trade-off is quality versus cost. Cheaper models lower the per-token rate but often raise the reject rate, so the retries can erase the savings. Before generating at full scale, run a small sample to measure your real reject percentage, then trim prompt length and tighten output schemas to reduce tokens per row. Small changes compound heavily across thousands of rows.
Rows times tokens per row times the model rate, plus the tokens burned on rejected generations you filter out. A cheap model at a dollar per million tokens makes tens of thousands of rows for a few dollars β the reject/retry rate is what moves the number.
Use a small, cheap model for bulk generation and reserve a stronger model for validation, tighten the prompt so fewer rows are rejected, and de-duplicate. The waste is in retries and near-duplicate rows, not the base generation.