How to Count Tokens and Estimate Your AI API Bill

To count tokens, run your text through the tokenizer of the model you will call. For OpenAI models that means the o200k_base or cl100k_base encoding in OpenAI's open-source tiktoken library. Anthropic and Google use their own tokenizers and offer count endpoints in their APIs. For a quick estimate, paste the text into the Token Counter.

To estimate the bill, count input and output tokens separately, multiply each by its price per million tokens and add them up. Output usually costs more, and on reasoning models it includes thinking you never see. The sections below show the code, the formula and a worked example with placeholder prices you replace with your provider's.

How many tokens your text is#

OpenAI and Google both publish a rule of thumb for English:

Source: OpenAI Help Center; Google Gemini API docs, checked October 2026

Rules of thumb break down outside plain English. Here is a 44-word support message in English and its translations into Spanish, German and Russian, counted with the two OpenAI encodings. In this sample, the newer o200k_base vocabulary needs far fewer tokens for the non-English versions.

Code behaves differently again: a 241-character Python function came to 58 tokens with both encodings, about 4.2 characters per token. Tokenizers also change between model generations. Anthropic's docs say that its newer tokenizer, used from Claude Opus 4.7 onward, produces roughly 30 percent more tokens for the same text than earlier Claude models, so do not reuse old counts after a model switch (Anthropic).

Count tokens for OpenAI models with tiktoken#

OpenAI publishes its tokenizers as tiktoken. Its model table maps model names to encodings. As of October 2026:

import tiktoken

enc = tiktoken.get_encoding("o200k_base")
text = "How many tokens is this sentence?"
print(len(enc.encode(text)))   # 7

tiktoken.encoding_for_model("gpt-4o") picks the encoding from a model name, but it raises an error for names its table does not know yet. In that case, choose the encoding yourself.

Two details make the API count a little higher than the text count:

• Message formatting. Each chat message adds a few tokens for roles and boundaries. OpenAI's cookbook uses 3 tokens per message plus 3 to prime the reply, and calls this "an estimate, not a timeless guarantee" (OpenAI Cookbook).
• Tools and files. Function definitions, images and files are counted on top of your text.

For an exact number that includes all of this, OpenAI's Responses API has a counting endpoint, POST /v1/responses/input_tokens, which takes the same input as a real request and returns input_tokens (OpenAI).

Count tokens for Claude and Gemini with their APIs#

tiktoken does not know other vendors' vocabularies, so its count is only a ballpark for Claude or Gemini. Both vendors offer an endpoint instead.

Claude. POST /v1/messages/count_tokens accepts the same body as a message request, including the system prompt, tools, images and PDFs, and returns input_tokens. Anthropic says it is "free to use" but rate limited by usage tier, and that the count is an estimate that "might differ by a small amount" from the real request (Anthropic).

import anthropic

client = anthropic.Anthropic()
count = client.messages.count_tokens(
    model="your-claude-model-id",
    system="You are a support assistant.",
    messages=[{"role": "user", "content": "Where is my invoice?"}],
)
print(count.input_tokens)

Gemini. The countTokens method returns the tokens in the input only (Google). With the Python SDK:

from google import genai

client = genai.Client()
result = client.models.count_tokens(
    model="your-gemini-model-id",
    contents="Where is my invoice?",
)
print(result.total_tokens)

Gemini counts media at fixed rates on its default settings: an image with both sides at 384 pixels or less is 258 tokens, larger images are split into 768×768 tiles of 258 tokens each, video is 263 tokens per second when processed at the default one frame per second, and audio is 32 tokens per second (Google). As of October 2026 the same page lists lower video rates for low resolution and for its agentic video mode, so check it for your model and settings.

Know what counts as input and what counts as output#

Your bill has two meters, and more goes into each than the text you type.

Input tokens are everything you send: the system prompt, the whole chat history, tool and function definitions, retrieved documents, images and files. In a chat, the history is resent with every turn.

Output tokens are everything the model generates: the visible answer plus any reasoning. OpenAI says reasoning tokens "are billed as output tokens" even though the API does not show them (OpenAI). Google labels its output price "including thinking tokens", and Anthropic bills thinking as output too.

Each API reports both meters in the response, so you can log real usage:

With Anthropic's prompt caching, input_tokens covers only the uncached part, so add the two cache fields to get the full input.

Estimate the bill with one formula#

Prices are quoted per million tokens, with separate rates for input and output. The formula:

cost = (input_tokens × input_price + output_tokens × output_price) ÷ 1,000,000

Here is a worked example for a support bot. The prices are placeholders, not any vendor's real prices: replace them with the current rates from your provider's pricing page.

Source: Worked example with placeholder prices of $1 and $5 per million tokens

Now change one assumption. If a reasoning model spends 800 hidden tokens thinking before each 400-token answer, output triples to 1,200 tokens per request. At the same placeholder prices, output costs $60 a day instead of $20, and the bill doubles to $80 a day.

Chat history grows the input side the same way. With an 800-token system prompt, 150-token questions and 300-token answers, turn n sends 800 + 150n + 300(n − 1) input tokens:

A ten-turn chat sends about 30,000 input tokens to produce 3,000 tokens of answers. This is also why long chats hit the context window; see tokens and context windows explained.

Reduce the tokens you pay for#

Start with whatever is sent on every request, because savings there multiply by your volume.

• Trim the system prompt and tool definitions. Cut examples the model no longer needs and tools it never calls in that flow.
• Cap the output. Set max_output_tokens (OpenAI Responses), max_tokens (Anthropic) or maxOutputTokens (Gemini), and ask for the length you need.
• Summarize or drop old history. Keep the last few turns verbatim and a short summary of the rest.
• Send excerpts, not whole documents. Retrieve the relevant passages; see what RAG is.
• Use prompt caching. Put the stable part of the prompt first. As of October 2026, OpenAI, Anthropic and Google all bill cached prefixes below the normal input rate; Anthropic, for example, charges cache reads at 0.1× the base input price on most models and cache writes at 1.25× or 2×, depending on how long the cache lives (Anthropic pricing). Multipliers vary by model, so check your provider's page.
• Batch work that can wait. OpenAI, Anthropic and Google each offer a batch API at 50% of the standard price, with results within a day (OpenAI, Google).
• Match the model to the step. Use a smaller model, or a lower reasoning effort or thinking budget, for classification, extraction and routing.

To count tokens, compare prices per model or track what each request costs, start with these:

P Prompt Token Counter

Monitor token usage for effective AI communication.

Free

💻 Developer tools

O OpenAI Cookbook

Resource for learning to use AI through practical examples.

Free + from $4.00/m

💻 Developer tools

L liteLLM

Access over 100 language models with ease and reliability.

Free

💻 Developer tools

O Openrouter.ai

Access multiple large language models for diverse AI tasks.

Free + from $0.50/m

💬 AI assistants

FAQ#