AI Token Cost Calculator

How the AI token cost calculator works

Every large language model API — OpenAI's GPT, Anthropic's Claude, Google's Gemini, xAI's Grok, DeepSeek and the rest — charges by the token. A token is a small piece of text, about four characters or three-quarters of a word. You pay for the tokens you send in (input) and the tokens the model writes back (output), and output is almost always the pricier side.

Cost per request = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)
Total cost = cost per request × number of requests

Providers quote prices per one million tokens, so this calculator does too. Pick a model preset to fill the rates automatically, or type your own to use it as an OpenAI API cost calculator, a ChatGPT API pricing calculator, a Claude or Gemini cost estimator, or a general LLM inference cost calculator. The comparison panel then prices your exact workload on every model at once — and you can price any other model, such as Meta's Llama, Alibaba's Qwen, Z.AI's GLM, Cohere, Perplexity or a Groq-hosted model, by typing its input and output rate into the custom fields.

How many tokens is that? Words and characters to tokens

For everyday English, one token is about four characters or roughly 0.75 words — so a word is about 1.33 tokens. Technical writing, code and non-English scripts usually use more tokens per word. Use these rules of thumb to size a prompt, or paste the text into the token counter above for an estimate.

2026 LLM API price comparison (per 1M tokens)

Current list prices, in US dollars, for 22 models across OpenAI, Anthropic, Google, xAI, DeepSeek and Mistral — verified against each provider's official pricing page in August 2026. This one table also powers the model dropdown and the live comparison above, so everything stays in sync. Prices change often and vary by context length, so confirm on the provider's page before you budget.

Estimates only — providers can change prices at any time and often launch new tiers. Cloud marketplaces such as AWS Bedrock and Azure OpenAI typically add roughly 10–30% on top of these direct-API rates.

GPT-4o vs Claude vs Gemini vs DeepSeek: which is cheapest?

There is no single winner — it depends on your token mix and how much quality you need. As a rough 2026 guide for a high-volume chatbot, the budget tiers win big: Gemini 2.5 Flash-Lite and DeepSeek are usually the cheapest, GPT-4o mini and Gemini Flash sit in the value zone, GPT-4o, Claude Sonnet and Gemini Pro are the mid-tier quality picks, and Claude Fable 5, Claude Opus and GPT-5.5 are the priciest per token. Instead of trusting a headline, enter your own input and output split above and read the Compare every model table — it ranks all 22 models cheapest-first for your exact workload, so the real answer is specific to you.

A worked example: monthly chatbot API cost

Imagine a support chatbot on GPT-4o handling 10,000 requests a month with about 1,000 input and 500 output tokens each. Input costs 10,000 × 1,000 ÷ 1,000,000 × $2.50 = $25; output costs 10,000 × 500 ÷ 1,000,000 × $10 = $50 — about $75 a month, or $0.0075 per request. The same workload on GPT-4o mini is roughly $4.50 a month, and Gemini 2.5 Flash-Lite is under $2 — which is exactly why the comparison table is worth a look before you pick a model.

Prompt caching and batch discounts

Two levers cut real money and are built into the calculator. Prompt caching stores a reused chunk — typically a fixed system prompt or shared context — and bills it at roughly 10% of the normal input rate on OpenAI, Anthropic and Google. For an agent that resends the same large instructions every call, caching can trim 40–70% off input cost; set your expected cached input % to model it. Batch APIs run non-urgent jobs (summaries, back-fills, evaluations) at about half price — tick Batch API to apply a 50% discount.

Reasoning (“thinking”) models cost more than they look

Reasoning models think in hidden tokens that never appear in the reply but are billed as output — often three to five times the visible answer. If you use a thinking or reasoning tier, budget your output tokens up, not down: a reply that shows 500 tokens might really bill 2,000–3,000. The calculator treats whatever you enter as billed output, so raise the output figure to stay realistic.

How much does an AI chatbot or agent cost to run?

Different workloads have very different token shapes. A chatbot is light — a few hundred tokens each way — so a small-business bot handling 100 conversations a day costs only a few dollars a month on a budget model, and even a frontier model stays modest. A RAG (retrieval) pipeline stuffs 5,000–10,000 input tokens of context per query, so input price and caching dominate. A coding agent reads whole files and reasons in long loops, burning hundreds of thousands of tokens in a single session — the fastest-growing and priciest pattern in 2026. To budget per user or per day, set the requests field to your conversations-per-day and switch the period to Day. Match the model to the shape: cheap tiers for high-volume chat, mid tiers with caching for RAG, and frontier models only for the hard reasoning steps.

Tips to lower your AI API bill

Most AI cost optimization comes from model choice and prompt discipline: use a smaller tier (GPT-4o mini, Claude Haiku, Gemini Flash or DeepSeek) for routine tasks, keep prompts tight, cap the maximum output length, cache repeated context, and route background jobs through the batch API. To control an enterprise AI budget, think in total cost of ownership — measure spend per feature and per user, not just per month — because one runaway agent loop can outspend everything else combined.

The AI token cost formula

No black box — here is exactly what the calculator does, so you can reproduce it in a spreadsheet.

See your cost in any currency

AI providers bill in US dollars, but you can switch the display to 14 currencies — including the euro, pound, Indian rupee, Canadian and Australian dollar, yen and more — to sense-check a budget in the money you actually think in. So you can check the OpenAI, Claude or Gemini API cost in rupees (₹) for a project in India, in pounds (£) in the UK, or in euros (€) across Europe. Conversions use an approximate rate and are for reference only; your invoice is charged in USD.

How to use it & key terms

Pick a model (or enter prices), set your input and output tokens and how often you call the API, then press Calculate for the cost per call, per day and per month.

Why a chat's cost grows with every turn

The API is stateless: the model keeps no memory between calls, so your app has to resend the entire conversation every time it asks for the next reply. That single fact quietly reshapes the bill. On the first turn you send a short system prompt and one question; by the tenth turn you are sending that system prompt plus the whole back-and-forth that came before, every earlier message counted again as fresh input tokens. The reply you see can stay a few sentences long, but the input riding beneath it grows with each exchange. In other words, the cost of a conversation is driven less by what the model says back and more by how much history you keep handing it.

This is why a real conversation costs more than a single exchange suggests. If each message and reply is a few hundred tokens, a ten-turn chat does not cost ten times the first turn — it costs more, because every later turn re-reads all the earlier ones. The closing turns of a long session are its most expensive part, and a few marathon threads can dominate a monthly total that is otherwise made up of short, cheap chats. To budget honestly, resist the urge to price one question and multiply. Instead, picture a typical conversation from greeting to resolution, add up the input across all of its turns, and enter that whole-session average as the input tokens above so the estimate reflects how people really use the assistant.

Once you can see the growth, the ways to tame it are clear. You rarely need to resend everything: trimming the thread to the last few turns, or replacing older messages with a short running summary, caps the input instead of letting it snowball. A lean system prompt helps as well, since that fixed block repeats on every single call. And setting a firm limit on how much history you carry protects you from the worst case, where one runaway session quietly outspends hundreds of ordinary ones. The same discipline keeps your forecast reliable: decide the conversation length you actually expect, then hold usage close to it.

• Early turns — cheap; there is little history to resend.
• Later turns — pricier; each one carries the whole thread as input.
• The fix — cap or summarise history, and keep the system prompt short.