Claude API Cost Calculator

Anthropic meters the Claude API per token in each direction, with no seat licence or minimum commitment, so spend is a direct function of prompt size, reply size, and request volume. The tier structure trades capability against price: the Opus tier sits at the top, the Sonnet tier is the general-purpose middle at rates such as 3 dollars per million input tokens and 15 per million output, and the Haiku tier is the low-latency, low-cost option. Output is consistently priced around five times input across tiers, reflecting that the prompt is consumed in one parallel prefill pass while every generated token requires a separate sequential pass. All three tiers expose a 200,000-token context window shared between prompt and completion. Prompt caching is the main structural lever on cost. A cached prefix is written once at roughly 1.25 times the normal input rate and then read at approximately one tenth of it, so the break-even is two requests against the same prefix under the default five-minute time-to-live; a one-hour cache costs about twice the input rate to write and needs three or more reads to pay off. The mechanism is a strict prefix match: a single changed byte before the breakpoint - a timestamp, a reordered JSON key, an added tool definition - invalidates everything after it, which is the usual reason a cache appears configured but never hits. Prefixes below roughly one to four thousand tokens do not cache at all. Work that tolerates latency can use the batch endpoint at a 50 percent discount.

Anthropic was founded in 2021 by Dario and Daniela Amodei together with a group of colleagues who had previously worked at OpenAI, with a stated focus on AI safety research. Its first assistant model, Claude, was released in March 2023, initially through partner access rather than a public self-serve API. Claude 2 followed in July 2023 and was notable for offering a 100,000-token context window at a time when most competing models were limited to a few thousand tokens, which made whole-document analysis a practical API use case and shifted the cost question from whether a document would fit to what it would cost to process. The three-tier naming that this calculator reflects arrived with the Claude 3 family in March 2024, when Haiku, Sonnet, and Opus were released together as speed-optimized, balanced, and capability-optimized options sharing a 200,000-token window. Claude 3.5 Sonnet followed in June 2024. Two pricing mechanisms were introduced later that year and changed how production systems were budgeted: prompt caching entered public beta in August 2024, letting applications that resend a large stable prefix pay a fraction of the input rate on repeat requests, and the Message Batches API arrived in October 2024 offering half-price asynchronous processing. Together they turned cost engineering for the Claude API into a question of prompt structure and scheduling rather than raw token volume alone.