Openai API Cost Calculator

OpenAI bills language model usage by the token rather than by the request or by elapsed compute time, so an application budget is governed by how much text moves through the model in each direction. A token is a sub-word unit produced by a byte-pair encoding tokenizer; for ordinary English prose one token averages roughly 4 characters or about 0.75 words, which puts 1,000 tokens near 750 words. Rates are quoted per million tokens and are always asymmetric, with output billed at 2 to 4 times the input rate because generating text requires one forward pass per token while the prompt can be processed in a single parallel prefill. On the rates modelled here, GPT-4o is 2.50 dollars per million input tokens against 10.00 for output, GPT-4o Mini is 0.15 and 0.60, and GPT-4 is 30.00 and 60.00 - a 200-fold spread between the cheapest and dearest input rate. The main budgeting trap is that input tokens are re-billed on every call. A chat application resends its system prompt, tool definitions, and the entire prior transcript with each turn, so cumulative input cost grows roughly with the square of conversation length rather than linearly, and retrieval pipelines add several thousand context tokens per request for the same reason. Reasoning models such as o1 and o3-mini bill their internal chain of thought as output tokens even though those tokens are never displayed, which can multiply the apparent cost of a short answer. Published rates change often, and discounts for cached input or asynchronous batch submission can cut the effective price materially, so any figure produced here is a planning estimate that should be reconciled against actual usage reporting.

Metered token pricing arrived with the OpenAI API itself, which opened as a private beta in June 2020 alongside GPT-3. That launch set the convention still in force: charge for the sum of prompt and completion tokens rather than for compute time or per query. The tokenizer behind the meter descends from byte-pair encoding, a compression scheme adapted to neural machine translation by Rico Sennrich, Barry Haddow and Alexandra Birch in 2016 and extended to a byte-level variant for GPT-2 in 2019. OpenAI later published that tokenizer as the open-source tiktoken library, which is why developers can count billable tokens locally before a request is ever sent. Rates were originally quoted per 1,000 tokens. The March 2023 release of gpt-3.5-turbo at 0.002 dollars per 1,000 tokens made a single chat turn cheap enough that per-1,000 figures grew awkward, and after GPT-4 Turbo was introduced at OpenAI DevDay in November 2023 the industry settled on per-million quotes. The gap between input and output pricing widened as models scaled: GPT-4 launched in March 2023 with output at twice the input rate, while GPT-4o in May 2024 set output at four times input. The o1 reasoning family, previewed in September 2024, added another layer by charging for hidden reasoning tokens, moving cost estimation away from counting visible words and toward modelling how much unseen work a model performs.