AI Token Counter & Cost Estimator

Why token count — not word count — determines what you pay

Every commercial LLM API bills by the token, and a token is not a word. It is a sub-word unit produced by a tokenizer: a byte-pair-encoding (BPE) vocabulary of a few hundred thousand entries that the model was trained against. The word unbelievable is one word but typically three tokens (un, believ, able). A single emoji can be two tokens. A two-character CJK phrase can be two or three. That is why the popular "divide characters by four" rule of thumb drifts by 20–30% on real text, and why an estimate that is good enough for a budget meeting is not good enough for a monthly invoice you have to defend.

This counter runs a BPE-style pre-tokenisation pass — the same word / number / punctuation splitting rules that GPT-family tokenizers apply — then estimates sub-word splits per chunk. On typical English prose it lands within about 10% of the exact count. Treat it as a planning number, not an invoice number.

How to get an exact count

There is only one authoritative source for the number you will be billed on, and it is always the provider:

• OpenAI — use the tiktoken library with the encoding matching your model, or read response.usage.prompt_tokens and completion_tokens from any API response.
• Anthropic — call the /v1/messages/count_tokens endpoint before sending, or read the usage object on the response.
• Google Gemini — use countTokens from the SDK, or the usageMetadata field on responses.

import tiktoken

enc = tiktoken.get_encoding("o200k_base")   # GPT-4o family
n = len(enc.encode("Your prompt goes here"))
print(f"{n} tokens")

The usage object on a real response is always preferable, because it also captures tokens you did not write yourself: tool schemas, system-injected preambles, and chat history that your framework appends silently.

The cost formula

Providers publish separate input and output rates per million tokens, and output tokens are consistently priced higher — usually 3× to 5× — because generation costs more compute than ingestion. Total cost is:

cost = (input_tokens  / 1_000_000) * input_price
     + (output_tokens / 1_000_000) * output_price

This matters more than most teams expect. A RAG application that stuffs 40,000 tokens of retrieved context into every request but returns a 200-token answer is almost entirely input cost. A code-generation endpoint that receives a 500-token prompt and emits 4,000 tokens of code is almost entirely output cost. The two workloads have opposite optimisation strategies, and knowing which one you have tells you where to spend engineering effort.

Four things that quietly inflate your bill

1. Tokenizer differences between providers

The same text produces different token counts on different providers. Anthropic's newer models use a tokenizer that can produce meaningfully more tokens for identical input than the GPT-family encodings. If you are comparing two vendors on published per-million prices alone, you can be misled: a 20% cheaper headline rate can be erased by a tokenizer that emits 25% more tokens. Always benchmark with your own real prompts.

2. Chat history re-sent on every turn

Most chat applications resend the entire conversation on each request. A 30-turn conversation does not cost 30× one turn — it costs roughly the sum of an arithmetic series, because turn 30 resends turns 1 through 29. Truncation, summarisation of older turns, or provider-side prompt caching are the standard fixes.

3. Prompt caching

Every major provider now discounts repeated prefixes. If your system prompt and tool schemas are identical across requests and you are not using caching, you are paying several times over for the same tokens. Cache-read discounts are steep — frequently an order of magnitude off the base input rate.

4. Reasoning models

Reasoning models emit internal reasoning tokens before answering. Those are billed as output tokens and are invisible in the final answer. A request that returns 300 visible tokens may bill for several thousand.

Privacy note

Everything on this page runs in JavaScript in your browser tab. Your text is never sent to a server, never logged, and never used to train anything. You can verify this by opening your browser's network inspector before typing — or by loading this page and then disconnecting from the internet entirely. The tool keeps working.

Frequently asked questions