Prompt Cost Estimator

Per-call API cost is the sum of two independently priced streams: input tokens covering the system prompt plus the user message, and output tokens covering the reply, each billed at a published rate per million tokens. Token counts are estimated here from word counts at about 1.3 to 1.35 tokens per English word, the ratio typical of byte-pair encoding tokenizers on ordinary prose. Output is nearly always the more expensive stream, commonly three to six times the input rate, and the reason is architectural: the prefill pass over the prompt processes all tokens in parallel and saturates the accelerator, whereas decoding emits one token at a time and is limited by memory bandwidth, so each generated token occupies far more hardware time than each ingested one. The practical consequence is that the two halves of a prompt reward different optimisations. A long system prompt is paid on every single request, so trimming it produces recurring savings, and that same repetition is why prompt caching, which discounts an unchanged prefix by 50 to 90 percent, exists at all. Output length is governed by instruction and by the max tokens parameter, which caps generation but is charged only for tokens actually produced. Caveats worth noting: word-based estimation drifts on code, JSON, and non-Latin scripts where token counts run considerably higher; published rates change often and vary by region and tier; and the figures here exclude cached-read pricing, batch discounts, and any per-request minimums a provider applies.

Metered language model access began with the OpenAI API in June 2020, which opened GPT-3 through a hosted endpoint and priced usage per thousand tokens rather than per request or per seat. The choice mattered because tokens, not words or characters, are what the model actually processes, so the billing unit matched the compute unit. Early tiers were named after Ada, Babbage, Curie, and Davinci, with the largest model listed at 0.06 dollars per thousand tokens, orders of magnitude above where comparable capability sits today. Separate input and output rates arrived with chat models. The ChatGPT API in March 2023 launched at a single low blended rate, but the GPT-4 release that same month priced prompt and completion tokens differently, formalising the asymmetry that mirrors prefill versus decode cost, and Anthropic and Google adopted the same structure for their own models. Two further mechanisms emerged in 2024 in response to observed production patterns: batch endpoints at roughly half price for latency-tolerant jobs, and prompt caching, offered by Anthropic and then OpenAI, which discounts repeated prefixes and was a direct answer to applications resending a large fixed system prompt on every call. Quoted rates also migrated from per-thousand to per-million tokens as unit prices fell.