Why are output tokens so expensive?
Output is generated sequentially, one forward pass per token, while input is processed in parallel. Providers price that difference at roughly 3–5× the input rate.
Estimate monthly LLM spend from request volume and token counts, compare model pricing, and see how caching or a smaller model changes the bill.
Cost per request = (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate). Monthly = cost per request × requests × (1 − cache hit rate on input).
50,000 requests a month with 2,000 input and 500 output tokens at $2.50/$10 per million costs about $500 — roughly half of it output tokens despite being a quarter of the volume.
Output is generated sequentially, one forward pass per token, while input is processed in parallel. Providers price that difference at roughly 3–5× the input rate.
Prompt caching on long system prompts (often 50–90% off cached input), capping max output tokens, and routing easy requests to a smaller model. Together these typically halve the bill.
Roughly 0.75 words per token in English, so 1,000 tokens is about 750 words. Code, JSON and non-Latin scripts consume noticeably more tokens per character.