Token cost calculator
Free, instant, no signup. Set the input, output and cached tokens per request and your monthly volume, and get per-request and monthly API cost across 22 Claude, GPT and Gemini models — each one next to its provider's list price. Rates are read live from the catalog.
Your workload
Rate: $1.20 in / $6.00 out · cache hit $0.1200 · cache write $1.20 — all per 1M tokens, read live from the Kunavo catalog.
Estimated cost
You save vs Anthropic list
$450/ month · ~60%
Estimate only. Reasoning models bill hidden reasoning tokens at the output rate, so a reasoning workload costs more than the output box suggests. Cache writes are billed separately at $1.20 per 1M and are not in this figure — they are a one-off per cache entry, not per request. Failed requests are never billed.
If you do not know your own token shape yet
Do not guess it from character counts. Log usage.prompt_tokens and usage.completion_tokens from a day of real traffic — every OpenAI-compatible response carries them — and put those numbers in above. Until then the presets encode realistic shapes: a support reply is light at roughly 800 in and 200 out, a RAG answer carries retrieved context at roughly 6,000 in and 500 out, one coding-agent step is heavier at 20,000-plus in, and long-document analysis is input-dominated.
Four levers that shrink the number
Prompt caching. A large stable system prompt or knowledge block bills at a fraction of the input rate once cached. For prompt-heavy applications this is the largest single lever there is — larger than switching models — which is why the calculator above has a cached-tokens field at all.
Model tiering. Route easy calls to the cheapest model that can do them and reserve the frontier tiers for the hard fraction. Set two models in the picker above and the delta is immediate.
Cap max_tokens. Output costs several times input across every family here, so unbounded generation is unbounded spend. This is the one control that needs no measurement first.
Trim retrieval. Fewer, better-ranked RAG chunks cut input tokens without hurting answer quality, and input is where a retrieval app spends.
Pricing guides for each model family
Cache hits at a tenth of input
Cached input and reasoning tokens
Haiku / Sonnet / Opus rates
Flash / Pro rates + free tier
GPT-5.4 / 5.5 / Mini rates
One API for every model
Per-image pricing
Every model, vs official
How do you calculate the cost of LLM tokens?
LLM API cost is priced per million tokens, with input and output billed at different rates, so the cost of one request is (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate). Output is typically several times more expensive than input, which is why a short prompt with a long answer can cost more than a long prompt with a short one. Monthly cost is that per-request figure multiplied by request volume, and cached input adds a third term at a fraction of the input rate.
Does this token cost calculator count tokens from text?
No — this calculator prices token counts you supply, it does not tokenize text. Enter the input, output and cached tokens per request and it returns per-request and monthly cost. To get real token counts from your own traffic, read them from the usage field your API responses already return rather than estimating from character counts, which drift by language and by tokenizer.
Is this token cost calculator free?
Yes. The token cost calculator is free, runs in the browser, and requires no signup or API key. Rates are read live from the Kunavo model catalog, and each model is shown next to its provider's published list price so the comparison is checkable against the provider's own pricing page.
Which LLM API is cheapest for my workload?
It depends on the input-to-output ratio, and that is the reason to compare rather than assume. A workload that ships long context and returns short answers is dominated by the input rate; a workload that generates long text is dominated by output, where the spread between tiers is widest. Set your own token counts above and read the monthly column — the cheapest model on a summarisation job is often not the cheapest on a chat job.
Why is my real API bill higher than the calculator says?
Three things a per-request estimate leaves out. Reasoning tokens are billed as output even though they never appear in the response. Cache writes are billed once per cache entry and sit outside the per-request figure. And retries are billed: a failed request that consumed tokens still counts. Agentic tools are the widest gap, because one user-visible task is many billed round trips.