Before an AI model like Claude can read what you wrote, it cuts your text into small pieces. Those pieces are tokens. They are the unit everything else is measured in: how much the model can hold, what it costs, and how fast it answers.
Try it
Watch a sentence get cut up
Type anything below. Each colored piece is one token. Common words are one piece. Long words, code, and other languages break into several. Colors only mark where one token ends and the next begins.
The program that does the cutting is called a tokenizer. Every AI company has its own, so counts differ a little between Claude, GPT, and Gemini. This page uses OpenAI's, because it can run right here in your browser.
Try:
Tokens
0
Words
0
Characters
0
Tokens per word
0
loading tokenizer…
Rule of thumb
A token is about ¾ of a word
For everyday English, one token is about 4 characters or ¾ of a word; that is Anthropic's own rule of thumb, and it says the count varies by language and content.[1] The other three figures below are that rule multiplied out.
1 token
≈ 4 characters, or ¾ of an English word
100 tokens
≈ 75 words, a long paragraph Worked out as 100 × ¾
1,300 tokens
≈ a 1,000-word essay Worked out as 1,000 ÷ ¾
100,000 tokens
≈ a 300-page novel Worked out as 300 pages × 250 words ÷ ¾
Why not just count words? Because the model never sees letters or words. It sees a list of numbers, one per token. Flip the switch above to see them. Rare words, names, and code get chopped into more pieces. Counts are approximate: every company slices a little differently, and even Claude's newest models slice into about 30% more pieces than older ones did.[1] The rule of thumb still holds.
Why it matters
Three things tokens decide
LIMITS
How much fits in the conversation
Every model has a maximum number of tokens it can hold at once: its context window. The newest models from Anthropic and OpenAI hold about a million tokens;[1][6] many everyday models hold 200,000. Your message, anything you attach, the whole conversation so far, and the reply all share that space. Fill it, and the model cannot take in any more.
00 used200,000
Tick what you would put into one conversation:
Item sizes are worked out from the ¾-word rule[1] at 250 words a page: a 20-page report ≈ 5,000 words; a 300-page novel ≈ 75,000 words; a year of email ≈ 450,000 words. This page was measured with the tokenizer above: about 4,000 tokens.
COST
What you pay for
AI services bill by the token: one price for what you send (input), a higher one for what the model writes back (output).[1] Prices are quoted per million tokens, so one turn of a conversation costs a fraction of a cent. It adds up at scale. The next section compares how Claude, GPT, and Gemini each do this.
Defaults are Claude Sonnet 5 list prices, $2 in and $10 out per million, as of September 13, 2026.[1] Edit them to match any model.
Input$0.0040
Output$0.0050
One turn$0.0090
× 1,000 turns$9.00
× 1,000,000 turns$9,000
SPEED
How fast the answer appears
The model does not write a reply and then show it to you. It produces one token at a time, each one chosen by looking at everything before it, and the app can show you each token the moment it is chosen.[4] That is why answers appear word by word, and why a long answer takes longer.
Press Play to watch a pre-written reply arrive one token at a time. Press again to skip to the end.
0 tokens0.0 s
Claude vs GPT vs Gemini
Same tokens, three different bills
All three companies charge per token, and all three split the bill into your message (input) and the reply (output). The differences are in the fine print: what counts as a reply, what a repeated message costs, and what happens when a message gets very long.
Newer models can also think before answering. That thinking is made of tokens you never see, and all three companies bill it at the reply price.[2][7][11]
Start by picking a situation. The three receipts below update.
Pick a situation:
Adjust the numbers yourself
This compares only the three models chosen above; change a dropdown and the answer changes. Cache-write fees and Gemini's hourly cache storage are not included.
Rule
Claude
GPT
Gemini
Two prices
Input and output, per million tokens. Opus 5 $5 / $25, Sonnet 5 $2 / $10, Haiku 4.5 $1 / $5.[1]
Input and output, per million tokens. GPT-5.6 Sol $4 / $20 (promo), Terra $2 / $12, Luna $0.20 / $1.20.[5][6]
Input and output, per million tokens. 3.1 Pro $2 / $12, 3.8 Flash $0.75 / $3.75 (intro).[10]
Thinking
Billed as output. The usage report tells you how many of the billed output tokens were internal reasoning.[2]
Billed as output, and not visible via the API; still takes up context space.[7]
Billed as output: “response pricing is the sum of output tokens and thinking tokens,” based on the full thoughts, not the summary you see.[11]
Repeated prompts
You mark what to cache (or turn on automatic caching). A cache hit costs 0.1× input. Writing costs 1.25× for a 5-minute cache or 2× for a 1-hour cache.[1]
On by default for prompts of 1,024+ tokens. Cached input costs 0.1×; on GPT-5.6 a cache write costs 1.25× and entries live 30 minutes after last use.[8]
Cached reads cost $0.20 vs $2 input on 3.1 Pro (0.1×), plus storage at $4.50 per million tokens per hour the cache is kept.[10] Minimum cacheable prompt is 2,048 to 4,096 tokens depending on model.[12]
Very long prompts
Flat rate across the full 1M window: “A 900k-token request is billed at the same per-token rate as a 9k-token request.”[1]
Above 272K input tokens, the whole request is billed at 2× input and 1.5× output (Sol: $8 / $30).[5][6]
Each company has its own tokenizer, so the same paragraph is a different number of tokens on each; Anthropic notes its own newer tokenizer yields about 30% more tokens than its previous one.[1] The counter at the top of this page uses the cl100k_base vocabulary, an OpenAI tokenizer.[14] Treat cross-provider counts as approximate.
The free option
Or run it yourself
The models on the receipts above only run in their companies' datacenters; you cannot download Claude, GPT, or Gemini. A different family, the open-weight models (Qwen, DeepSeek, GLM and others), can be downloaded and run on your own computer. Then nobody charges you per token. You pay for the machine and the electricity, and every token after that is free.
FreeToken is the newest and the most ambitious. Its trick: a Mixture-of-Experts model only wakes up a small slice of itself for each token, so FreeToken streams the slices it needs between your SSD, your RAM, and your graphics card fast enough to run a 284-billion-parameter model on a gaming PC.[15] What the authors measured, on the machines they name:[15]
Qwen3.6-35B, RTX 5090, reading a 16K prompt6,700 tokens/s
The graphics card in those tests, list price$1,999[16]
Could a heavy user go free?
People who hear that someone pays hundreds of dollars a month in tokens ask whether a free local model could replace it. Set the workload and see.
Pay per token
$252a month at list price, before caching discounts
Run it yourself
3.4 hof the RTX 5090 rig busy, every day
Local speeds are the paper's DeepSeek-V4-Flash figures on an RTX 5090 with 180 GB of RAM: 25 tokens/s writing, 6,700 tokens/s reading.[15] The machine itself is the $1,999 card plus a computer with about 180 GB of RAM.
A real example
My own use, counted from the files
This site is built with Claude Code, GPT Pro, and Gemini Pro, in many sessions at once. The Claude Code sessions write a transcript on my machine, and every reply in it carries the exact token counts the model billed, so those can be counted; the GPT and Gemini work is not in these files and is on top of everything below. The chart is the Claude counts, each billed reply once, added up per day.
Tokens per day, everything the models read and wrote
Almost all of it is re-reading: the same conversation history goes back to the model on every turn, from the cache.
Reply tokens per day, the part that is new text
What the model actually wrote, the tokens billed at the reply price.
So can it be free? The tokens can be. The machine cannot, the model changes, and the queue is real. Nobody gives away millions of frontier-model tokens a day. The three ways to get them are: pay per token, as above; pay for a subscription that comes with usage limits; or buy the hardware and run a smaller open-weight model, one session at a time.
One thing that trips people up
Not those tokens
The word gets reused all over tech. These are unrelated. In an AI context, a token is only ever a chunk of text.
Login tokens
A temporary key a website hands your browser to prove you have signed in. Security, not language.
Crypto tokens
A tradable digital asset on a blockchain. Nothing to do with text.
Arcade tokens
The coin you feed a machine. Closest in spirit, since you do spend AI tokens, but still not the same thing.
Sources
Where every number comes from
Each bracketed number on this page links here. All pages were read on September 13, 2026. Quotes are verbatim from the source.
Model table: Opus 5 $5 / $25, Sonnet 5 $2 / $10, Haiku 4.5 $1 / $5 per MTok; cache hits $0.50 / $0.20 / $0.10. 5-minute cache write 1.25x base input price … 1-hour cache write 2x … Cache read (hit) 0.1x.The Batch API allows asynchronous processing of large volumes of requests with a 50% discount on both input and output tokens.Claude 4.6 and later models … include the full 1M token context window at standard pricing. (A 900k-token request is billed at the same per-token rate as a 9k-token request.)As a rough estimate, 1 token is approximately 4 characters or 0.75 words in English. The exact count varies by language and content type.Claude 4.7 and later models … use a newer tokenizer … This tokenizer produces approximately 30% more tokens for the same text.
monitor the usage.output_tokens_details.thinking_tokens field in the response, which reports how many of the billed output tokens were internal reasoning.
GPT-5.6 Sol: input $4.00, cached $0.40, output $20.00; long context $8.00 / $0.80 / $30.00. Terra $2.00 / $0.20 / $12.00; long context $4.00 / $0.40 / $18.00. Luna $0.20 / $0.02 / $1.20; long context $0.40 / $0.04 / $1.80. GPT-5.5 $5.00 / $0.50 / $30.00. GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026. Batch pricing listed at 50% of standard rates.
Prompt caching is enabled by default for supported OpenAI models. Minimum 1,024 tokens for GPT-5.6 and later; cached tokens at 0.1× the uncached input-token rate; on GPT-5.6 cache writes at 1.25× the uncached input-token rate; entries kept 30 minutes after its most recent write or reuse.
When thinking is turned on, response pricing is the sum of output tokens and thinking tokens.Pricing is based on the full thought tokens the model needs to generate, despite only the summary being output from the API.
Batch API usage is priced at 50% of the standard interactive API cost for the equivalent model.Batch jobs are designed to complete within a 24-hour turnaround time.
gpt-tokenizer (open-source JavaScript port of OpenAI's tiktoken), cl100k_base vocabulary
The live splitter on this page loads gpt-tokenizer@2.9.0/dist/cl100k_base.js from jsDelivr and runs it in your browser; token ids shown are that vocabulary's ids.
FlashML, “FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution” (arXiv, August 2026)
Models: Qwen3.6-35B-A3B (35B, BF16 and NVFP4); DeepSeek-V4-Flash (284B, 13B active, MXFP4); GLM-5.2 (753B, 40B active, NVFP4). Hardware named per run: RTX 5090 (32 GB) with 180 GB DDR5; RTX 4060 Laptop (8 GB) with 32 GB; RTX PRO 6000 (96 GB) with 512 GB. Decode: Qwen3.6-35B 77–83 tok/s on the 5090, 39.3 tok/s on the 4060 laptop; DeepSeek-V4-Flash 22–25 tok/s on the 5090; GLM-5.2 14.9 tok/s on the RTX PRO 6000. Prefill: 6.7k tok/s at 16k tokens on the 5090 for Qwen3.6-35B.