AI Token Calculator for OpenAI, Claude & Gemini

What is an AI token estimator?

An AI token calculator estimates how many tokens a prompt, response, document excerpt, or chat history may use before you send it to a model. That makes it useful for early planning: you can measure prompt size here, then move the result into the AI API Cost Calculator to estimate request-level and monthly spend.

Tokens are the small units AI models read and generate. A token can be a character, a short word, part of a longer word, punctuation, or whitespace. API providers meter tokens because token volume is a closer proxy for model work, so token volume is the starting point for any AI budget.

Working with a spreadsheet or many prompts? Count tokens in a CSV or batch file.

How to calculate tokens for OpenAI, Claude, and Gemini

Start by separating the content you send from the content the model generates. For OpenAI, Claude, Gemini, and other APIs, input tokens usually include system instructions, user messages, chat history, retrieved context, and uploaded text. Output tokens are the answer the model writes back. Provider tokenizers can differ, so treat this page as a fast planning estimate and verify production-sensitive workloads against official tools and pricing pages.

For provider-specific cost context, compare the current model list in the AI Model Pricing Comparison, review OpenAI API pricing, or check Anthropic Claude API pricing before making a launch budget.

Token count vs API cost

Token count tells you the size of the work. API cost depends on the model price, how many input tokens you send, how many output tokens the model generates, and how often the workflow runs. A short prompt can still become expensive if it produces long answers at high volume, while a longer prompt may be acceptable when it improves accuracy and reduces retries.

For OpenAI API cost planning, measure the prompt, retrieved context, and expected answer here first. Then check whether current prices changed in the OpenAI API pricing change guide and move the token counts into the AI API Cost Calculator to estimate monthly spend.

Once you know the rough token range, use the Prompt Cost Optimizer to trim repeated wording and oversized context before those tokens become recurring spend.

• Token usage: measure the prompt, chat history, retrieved context, and expected response size before production traffic.
• Output costs: watch long answers, JSON responses, summaries, and multi-step agent outputs because generated tokens can be priced higher.
• API spend: move realistic token ranges into the API cost estimator to estimate cost per request, daily usage, and monthly API cost.

Words to tokens and tokens to dollars: quick reference

These tables use the same English planning ratio as the calculator's word mode: about 0.75 words per token, which sits between the 0.81 we measured for everyday English and the 0.66 for technical docs. Do not convert by characters: the common "one token is four characters" rule overcounts plain English and misses about half the tokens in JSON and CSV. Treat these tables as budgeting ranges, not exact tokenizer output; paste real text into the calculator above when the count matters.

Before you budget an agent that reads JSON or CSV, check the measured characters-per-token ratios for 11 kinds of text, including code and Chinese.

What tokens cost in dollars

API prices are quoted per 1 million tokens, with output usually priced higher than input. The table below converts current standard list prices for the lowest-priced model in the calculator and the two calculator defaults; compare every model in the AI Model Pricing Comparison.

How this estimate works

The calculator counts words, digits, punctuation, and CJK characters separately and gives each its own token rate. We measured those rates against OpenAI's o200k_base tokenizer on English prose, code, CSV and JSON tables, and Chinese text; the total estimate for each type landed within about 5% of the real count. A flat 4-characters-per-token rule counted number-heavy CSV at about half its real size. Claude and Gemini use their own tokenizers, so treat the result as a planning number, not a bill.

Worked examples

Example: 1,000 words to tokens

A 1,000-word English article came to about 1,240 tokens on OpenAI's o200k_base tokenizer in our measurement, and dense technical writing to about 1,530; headings, lists, numbers and code push it higher. If the same article is sent as retrieved context on every request, the token cost repeats every time the workflow runs. For budgeting, test a low, typical, and high-length example rather than relying on one clean sample.

Example: monthly chatbot token usage

Suppose a support chatbot sends a 700-token instruction, 1,200 tokens of retrieved help content, a 150-token customer message, and receives a 450-token answer. That is roughly 2,050 input tokens and 450 output tokens per reply. At 20,000 replies per month, the workflow uses about 41 million input tokens and 9 million output tokens before retries, escalation flows, or longer conversations are included.

Paste real prompts, retrieved snippets, onboarding messages, or customer support examples. Save typical low, medium, and high token counts, then use those ranges in the API Cost Calculator and AI Budget Planner instead of relying on a single best-case prompt.

How many tokens does a knowledge assistant for 5,000 employees use?

Short answer: about 550 million input tokens and 50 million output tokens a month in a typical rollout, and anywhere from roughly 140M to 1.8B input tokens depending on how many people actually use it. Usage, not head count, drives the number, so estimate it in four steps:

1. Active users: not every employee asks questions every day. Start with 15% to 50% of head count on a workday.
2. Questions per user: 2 to 8 a day is a common range for an internal help or policy assistant.
3. Input per question: 800 tokens of system prompt, 3,000 tokens of retrieved document chunks (six chunks of about 500 tokens), 500 tokens of recent chat history and a 100-token question, so 4,400 input tokens. Paste one of your real retrieved chunks into the calculator above to check the 3,000 figure.
4. Output per answer: about 400 tokens, or roughly 300 words.

Input is about 90% of the tokens here, so the cheapest lever is the context you send, not the answer length. The system prompt is identical on every request, which makes it a good fit for prompt caching; check the discount for your model in the Prompt Caching Savings Calculator. Retrieving 3 chunks instead of 6 cuts input by about a third. To turn these numbers into a monthly budget with your own model and traffic, use the AI API Cost Calculator, and set team or department limits in the AI Budget Planner.