Frequently asked questions

Less than most people expect, and the spread between choices is enormous. A support bot serving 5,000 people who ask about eight questions each runs $5.04 a month on the cheapest model we price, $52.20 on the median one, and $1,080 on the most expensive. Same bot, same users, same questions. The model you choose is the largest single lever, and routine support work rarely needs the top tier. Price your own version rather than trusting a range.

Work out one request, then multiply. Estimate what goes in (your instructions plus any documents you attach) and what comes back, price that on a mid-tier model, and multiply by the requests you expect in a month. Add margin, because a first version is always more talkative than the one you ship. For most small business projects the API bill lands in tens of dollars a month rather than thousands. The estimator does this from plain answers if you would rather not count anything.

Yes, close to linearly, which is the good news: ten times the users costs about ten times as much, with no seat licence and no step change. Two things bend that line. Conversation history grows if you resend it every turn, so a busy chat can cost more per user in month six than in month one. And prompt caching gets better with scale, because the repeated part of your prompt is reused more often. Model your own growth in the advanced calculator.

Quote the build once and the running cost separately, because they behave differently: your build fee is finished when the work is, and the API cost recurs forever and rises with use. A workable structure is a fixed build fee, then the API cost passed through monthly with a cap you both agree. Set the cap using the flagship figure from the estimator rather than the one you expect to pay, so a busy month cannot cost you money. The trap is quoting one fixed price that swallows an unbounded variable cost, which turns your most successful client into your least profitable one.

Usually not, and you should plan around that. At small business scale the API bill is often the smallest line. The costs that actually dominate are the build itself, connecting it to the systems you already run, the testing that tells you it is safe to trust, and the human time spent checking its output for the first few months. Budget for a person reviewing the output early on, because that is the line most projects forget and the one that decides whether the thing survives contact with real customers. The API figure is worth knowing because it is invisible in advance and it never stops, but it is one input to the decision rather than the decision.

No. Tell the estimator what you are building and roughly how much of it you do, and it returns a monthly figure. Tokens are just the unit providers bill in, the way a phone plan bills minutes. If you want the detail it is all in the breakdown, and the token counter shows exactly how your own text is measured, but you can get a usable number without opening either.

For English text, one token is roughly 4 characters or about 0.75 words, so 1,000 tokens is around 750 words. Code and non-English languages usually consume more tokens per word. This calculator uses the 4 characters per token heuristic for instant client-side estimates.

Generating output tokens requires a full forward pass of the model per token, while input tokens are processed in parallel. Output tokens are therefore more computationally expensive, and most providers price them well above input, commonly several times higher.

Prompt caching lets a provider store the processed state of a repeated prompt prefix, such as a system prompt or shared documents, and charge a heavily discounted rate when it is reused. Cached input tokens typically cost 75% to 90% less than fresh input tokens, so workloads with large repeated prefixes can cut input costs by up to 90%.

Yes. OpenAI, Anthropic and Google all offer batch endpoints that process requests asynchronously, usually within 24 hours, at roughly a 50% discount on both input and output tokens. Batch pricing is ideal for evaluations, backfills, and any job that does not need a real-time response.

Only at sustained high volume. A single 80GB GPU costs roughly $1,200 to $2,500 per month in the cloud before engineering time. Serverless API pricing for open models like Llama is so competitive that most teams need millions of requests per month with steady utilization before self-hosting wins.

Prices are captured nightly from a continuously maintained public pricing dataset (LiteLLM) into our own archive and served from our own domain, with the dataset's public copy and a built-in snapshot as fallbacks. Estimates are for planning purposes; always confirm current rates on the provider's official pricing page before committing to a budget.

It runs a free optimization pass entirely in your browser: stripping conversational filler and redundant whitespace, removing duplicate context blocks, and reordering the prompt so static instructions come first and dynamic variables last, which is the structure prompt caching discounts require. You get a before and after comparison with token counts, projected dollar savings at your scale, and suggestions for cheaper models that can handle the cleaned-up workload. Your prompt never leaves your device.

Yes. The calculator and the prompt optimizer are 100% client-side JavaScript: text you paste is processed on your device and is never uploaded, logged, or stored on any server. The only network traffic is inbound pricing data, and you can verify this yourself by watching the network tab in your browser's developer tools while you use the tool. The sole exception is an email address you choose to submit for reports or alerts.

It changes as providers reprice, which is why this page syncs live pricing data on every load. As a general rule the budget tier (flash and mini class models, DeepSeek's chat models) runs 10x to 30x cheaper than flagships. The cheapest choice for your specific workload depends on its input/output shape and caching, so paste your prompt above and the top row of the ranking is your live answer.

Multiply your monthly calls by the cost per call: input tokens times the input rate plus output tokens times the output rate (rates are per million tokens). Apply any caching discount to the repeated part of your input. The calculator above automates exactly this math across every major provider at once.

No. The prices shown are standard on-demand rates. Batch endpoints from OpenAI, Anthropic and Google run about 50% cheaper for asynchronous jobs, and fine-tuned models carry their own per-token rates. Treat the calculator's results as your on-demand baseline and apply batch discounts to any deferrable share of the workload.

Yes. The counter is provider-agnostic: it uses the 4 characters per token heuristic, which lands within roughly 10 to 15% of the exact counts produced by OpenAI, Anthropic, and Google tokenizers for English text. For billing-grade precision use the provider's own tokenizer; for budgeting and comparison, the estimate here is plenty.