Use this free AI token cost calculator to separate uncached input, cached input, and output charges without relying on stale model-price presets.
10000 requests; no cached input included
No cached input is included in this result. Add the cached part of total input and the current cache-read rate when your provider usage log reports cache hits. Cache writes, storage, batch rates, tool calls, retries, credits, and taxes still need separate checks.
Recent token cost estimates will appear here.
Token cost math stays local. Enter current standard, cached-input, and output rates because provider pricing can change.
Inputs and recent answers stay in this browser tab and are not sent to a server.
Estimate the monthly cost of an AI support bot, writing helper, or internal tool.
Measure how much a repeated prompt prefix could change input cost when cache reads are reported.
Compare two model price cards using the same token and request assumptions.
Turn a token estimate into a rough budget before building a prototype.
Explain why long prompts and long answers can cost different amounts.
$8 uncached input + $1.60 cached input + $40 output = $49.60 total, or $0.00496 per request.
$0.045 input + $0.09 output = $0.135 total, about $0.000135 per request.
$20 input + $7 output = $27 total; long prompts drive most of the cost.
Need a slower walkthrough, a related calculator, or the full library? These links keep you close to the task you started.
Plain-language answers about when to use the tool, what it does with your inputs, what to double-check, and how privacy works.
Use it when your task matches one of these common needs: Estimate the monthly cost of an AI support bot, writing helper, or internal tool. Measure how much a repeated prompt prefix could change input cost when cache reads are reported. It works best when you already know the measurements, amounts, units, or options the page asks for.
In plain language: Uncached input tokens per request = total input tokens - cached input tokens. Uncached input cost = uncached input tokens * requests / 1,000,000 * standard input price. Cached input cost uses the cached-input tokens and cached-input price. Output cost uses output tokens and output price. Total cost = uncached input cost + cached input cost + output cost. The examples on the page are there so you can compare your inputs with a worked example before copying the answer.
Requests: How many model calls you want to estimate, such as one day, one month, or one product test. Total input tokens per request: All tokens sent to the model, including instructions, prompt text, context, tool messages, and any cached part reported inside that total. Cached input tokens per request: The part of total input served from cache. Enter 0 when you have no cache-read count. Output tokens per request: Tokens generated by the model in each response. Standard input price per 1M tokens: The current uncached input rate for the model and service tier you plan to use. Cached input price per 1M tokens: The current cache-read rate. Cache creation or storage can be a separate charge. Output price per 1M tokens: The current provider rate for one million output tokens. This is often different from the input price.
This is a planning estimate, not a live provider bill. Cached input must be part of the total input count, not extra input added on top. Providers can charge separately for cache writes, cache storage, long context, service tiers, batch jobs, tools, images, audio, credits, and taxes. Use the current rate card and real usage logs for budgets that matter. Also check the unit, scale, mode, and result limit because small input changes can change the answer.
Start with total cost, then compare uncached input, cached input, and output cost. Cost per request and cost per 1,000 requests make the scale easier to judge. Requests for $100 is a runway estimate using the same average token counts and rates, not a provider spending limit.
Model prices change and different providers charge different rates for input, output, cached input, fine-tuned models, batch jobs, and special tools. Entering the rate yourself keeps the calculator useful without pretending one price is always current.
Enter cached input as the part of total input that received the cache-read rate. The calculator subtracts that part from total input, prices the uncached remainder at the standard rate, and prices cached input separately. Do not add cached input on top of total input.
Use the provider's current batch input, cached-input, and output rates in the price fields. Do not apply a guessed universal discount because model support, service tiers, cache treatment, and batch rules can change.
Use the number of requests you expect in a month, then enter the average input tokens and output tokens for one request. The total cost is the input spend plus output spend for that whole request count.
Many model providers charge different rates for tokens you send and tokens the model generates. Long prompts raise input cost, while long answers raise output cost, so keeping them separate makes the estimate easier to check.
Only if your token numbers came from the exact tokenizer or usage logs for the model. Rough text estimates can be useful for planning, but code, symbols, non-English text, whitespace, and tool messages can change the real token count.
Retries, longer system prompts, tool messages, retrieval context, cache writes, cache storage, long-context tiers, and streamed follow-ups can change the bill. Pull input, cached-input, and output averages from provider logs when the app has traffic.
Yes. Keep the request count and token assumptions the same, then enter one model price card and compare it with another. This shows the pricing effect, not quality, latency, rate limits, or reliability.
No. It only estimates model token charges from the rates you enter. Add hosting, storage, retrieval, image/audio/video tools, retries, monitoring, and other platform costs separately.
No. The calculator runs in your browser tab. Your recent answers stay only on the page while you use it, and they are not sent to a server.