LLM API Cost Calculator

Input and output tokens are normally priced separately. Input includes the instructions, conversation history, retrieved documents, tool descriptions, and other context sent with a request. Output is what the model generates. A request count on its own is therefore not a useful cost unit: two endpoints can receive the same number of calls while one repeatedly sends a long document and the other sends a short classification prompt.

Use provider usage fields or production telemetry to calculate a representative input-token and output-token distribution. An average is useful for a first budget, but a percentile view is safer for products with occasional large contexts. Record whether the observed token count already includes retries, follow-up tool calls, and hidden orchestration steps. Counting only the user's first request can materially understate an agentic workflow.

For each priced token category, monthly cost equals monthly tokens divided by the provider's pricing unit, multiplied by the quoted price. If pricing is stated per one million tokens, 100 million input tokens at 2 currency units per million costs 200 currency units. Perform the input and output calculations separately, then add any other metered services that the calculator does not model.

Keep units consistent. A field labeled tokens per request must not receive characters, words, or bytes. Likewise, a price per million tokens must not be entered as a price per thousand. When comparing providers, use the same workload sample and the tokenizer or reported usage from each provider; the same text need not produce an identical token count everywhere.

Suppose an application handles 100,000 calls in a month, averaging 1,000 input tokens and 400 output tokens. That is 100 million input tokens and 40 million output tokens. With illustrative prices of USD 2 per million input tokens and USD 8 per million output tokens, the modeled cost is USD 200 for input plus USD 320 for output, or USD 520 in total.

The useful conclusion is not that the application will always cost 520. It is that output represents 320 of the modeled total even though output volume is lower. A team considering shorter responses can now estimate the budget effect without pretending that shortening output is free from quality trade-offs. If 5% of calls are retried and those retries are absent from the traffic input, a separate retry scenario should be added rather than hidden in a rounding buffer.

Official provider pages may publish different prices for cached input, cache creation, batch execution, priority processing, or longer contexts. Those categories are not interchangeable. A cache discount applies only when the provider records a qualifying cache hit; a repeated prompt in application logs is not proof that every token received the cached rate. Batch pricing may also come with different latency or processing conditions.

Tool-enabled requests can create charges outside the visible text-token line. Examples include web search, file retrieval, image or audio processing, code execution, storage, and repeated model calls made by an orchestration loop. Build a bill of materials around the workflow. Use this calculator for the token categories it exposes, then add provider-specific metered features as separately sourced lines rather than forcing them into a token price.

A defensible comparison replays the same representative workload, measures usage, and checks whether each setup meets the application's accuracy, latency, safety, and reliability requirements. The lowest modeled token bill is not automatically the lowest total cost if it requires more retries, longer prompts, additional validation, or a larger operational burden. Conversely, an expensive price per token can still be economical for a low-volume step that removes substantial manual work.

The calculator produces a deterministic estimate from entered traffic and token assumptions plus the visible source-dated rate presets. It does not query provider billing APIs, predict tokenization, model tiered contract discounts, convert currencies, apply tax, value service credits, or determine whether a workload qualifies for cache or batch pricing. It also cannot estimate the business cost of latency, an incorrect answer, abuse, or a production outage.

Treat the result as a budget bridge between product telemetry and an official rate card. Reconcile it against actual provider invoices after launch. If the difference is material, investigate uncounted endpoints, retries, long-context requests, non-text tools, and rate-category assumptions before merely increasing the forecast.

Provider models, tiers, and prices can change. This guide and the visible standard-price presets were fact-checked on 20 July 2026, but that date is not a price guarantee. Open the linked official pricing page immediately before a purchasing or architecture decision, confirm the effective model and price categories, and save the review date with the scenario.

Refresh usage assumptions after prompt changes, model migrations, retrieval changes, or meaningful traffic shifts. A monthly review is appropriate during launch; a stable service can move to a longer cadence while still reconciling estimates with invoices. A copied price without its model, unit, and review date is not an auditable input.