What a Token Actually Is, and Why Your Bill Counts Them

Описание: Building with AI and want people to build with? Join the community: https://www.skool.com/ai-operators-4364

Your AI bill is not counted in words, characters or requests. It is counted in tokens. This video shows what a token is, with real text run through a real token counter, and where the tokens you pay for actually come from.
What we counted (Gemini's tokenizer, a free count endpoint, no model run): one long number is 1 word and 53 tokens. A 10 word English sentence is 12 tokens, the same sentence is 17 in Turkish and 16 in Japanese. The same JSON record is 154 tokens pretty-printed and 84 minified. A 23 token question becomes 570 tokens once a system prompt and five tool definitions ride along. A blank test image counted as 1,090 tokens. A nine message chat of 304 tokens was billed 948 input tokens over 5 requests (3.1 times), because every request sends the history again.
The price side: on the page we read, Google lists gemini-3.8-flash at $0.75 input and $3.75 output per 1M tokens, 5 times more for output. Anthropic's Claude Opus 5.5 and OpenAI's gpt-6.1-sol show the same multiple. Other tokenizers give other counts: count on the model you pay for.

FOUR MOVES (our opinion)
1. Count, do not guess. Anthropic and Google both document a token counting endpoint and both say counting is not billed. Count on the model you will run.
2. Shrink the boring parts. Minify JSON and send short IDs. Minifying alone was 45 percent fewer tokens in our test.
3. Audit what rides along. Every tool definition and every line of system prompt is paid for on every request. Send only the tools the task needs.
4. Watch the output. It is the expensive side. Ask for short answers when short is enough, and check how much thinking you pay for.

CHAPTERS
00:00 One number, 53 tokens
00:44 What a token is
01:34 We counted six texts
03:46 Input, output and the price
05:08 Tokens you never see
06:34 The chat that tripled
07:53 The context window
08:23 Four moves

SOURCES (all read 2026-10-13)
OpenAI Help Center, Understanding and counting tokens: https://help.openai.com/en/articles/4...
Google, Gemini API docs: Tokens: https://ai.google.dev/gemini-api/docs...
Anthropic, Claude docs: Token counting: https://platform.claude.com/docs/en/b...
Google, Gemini API docs: Billing: https://ai.google.dev/gemini-api/docs...
Anthropic, Claude docs: Glossary: https://platform.claude.com/docs/en/a...
Anthropic, Claude docs: Pricing: https://platform.claude.com/docs/en/a...
OpenAI, API docs: Pricing: https://developers.openai.com/api/doc...
Google, Gemini API docs: Pricing: https://ai.google.dev/gemini-api/docs...
OpenAI, API docs: Latency optimization: https://developers.openai.com/api/doc...
OpenAI, API docs: Reasoning models: https://developers.openai.com/api/doc...
Anthropic, Claude docs: Context windows: https://platform.claude.com/docs/en/b...
OpenAI, API docs: Function calling: https://developers.openai.com/api/doc...
Google, Gemini API docs: Media resolution: https://ai.google.dev/gemini-api/docs...
Anthropic, Claude docs: Vision: https://platform.claude.com/docs/en/b...
Prices and docs change. Read the page again on the day you decide, then count.

ON-SCREEN STAMPS
VENDOR'S DOCS: from that vendor's own page, read on 2026-10-13; a script re-checks every quote.
OUR RUN: a line of our own log: real text sent to Google's token counter for Gemini (a count endpoint, no model run).
OUR MATH: arithmetic on sourced numbers, logged and reproduced by a second script.
OUR OPINION: our judgment.
NOT STATED: no page we read gives it.
ILLUSTRATIVE: a made-up text or a drawing aid, not data.

Narration is AI-generated.
Not affiliated with Google, Anthropic or OpenAI. Company and model names are used as text only, to say what their pages say.
General information, not professional advice. The texts we counted are made up; the counts are real.