Same JSON extraction job costs $667 or $1,100 per 100,000 tasks, depending on validation rate

The price per token stays fixed, but the effective cost of a task depends on how often the model's output passes validation. An author calculates a support-ticket extraction at about $0.0066 per call on GPT-5.6, then divides by the share of valid answers to estimate the cost of a finished task. Higher temperature makes malformed JSON more likely, and each failed answer triggers a billed retry. Illustrative validation rates of 0.99 to 0.60 produce totals from roughly $667 to $1,100 for 100,000 tasks. The author notes these rates are assumptions rather than measurements and suggests logging failures by temperature to find one's own rate. Retries that reuse the same seed, temperature and prompt may simply repeat the failure, so they should change something and be capped.