Token Calculator
Paste your text to estimate token count and see what it would cost across all major AI models.
0
Tokens (est.)
0
Words
0
Characters
Token count is an estimate based on ~4 characters per token (English average). Actual tokenization varies by model and language. For exact counts, use each provider's tokenizer.
How token counting works
Every request to an LLM API is billed in tokens, not characters or words — a token is roughly 4 characters of English text, though this varies by language and by each provider's specific tokenizer. Paste any text above and we estimate its token count, then show what it would cost as either an input or an output across every major model, so you can see the price difference just from how the same content is used. This is useful for sizing prompts, estimating summarization costs, or sanity-checking why a bill came in higher than expected.
What actually eats your token budget
Prose is rarely the problem. The heavy consumers are structure and repetition: JSON keys, quotes and braces are all tokens, so a verbose API response schema can double the billed size of the same information. Code is denser still — indentation, symbols and camelCase identifiers split into more tokens per visible character than plain English does.
Language matters more than most teams expect. The common tokenizers were trained mostly on English, so Portuguese, Spanish or German text typically consumes 20-40% more tokens for the same content, and languages with non-Latin scripts can use several times more. If your product serves users in Portuguese, budget for that overhead — the same prompt template literally costs more per request than its English original.
Rules of thumb worth memorizing
A handful of conversions cover most day-to-day estimates. One token is about four characters of English, so 100 tokens is roughly 75 words. A dense page of text runs around 500 tokens; a 300-page book sits in the neighborhood of 150,000 — which is why whole books only fit in the largest context windows (see the context window calculator). Reading these numbers the other way is just as useful: a model that replies with a 400-word answer generated about 530 output tokens, billed at the expensive rate. Estimates first, tokenizer second: when a number actually matters for a contract or a budget, verify it with the provider's official tokenizer.
Five ways to shrink token usage
First, tighten the system prompt — it is resent with every request, so every trimmed sentence pays back on all future traffic. Second, ask for short, structured answers: a strict output format with explicit length limits routinely cuts output tokens in half. Third, strip boilerplate before sending — HTML tags, headers, footers and navigation text in scraped pages are billed like content. Fourth, summarize or truncate conversation history instead of replaying it in full. Fifth, match the model to the job: pricing per token differs by an order of magnitude across models, and the OpenAI cost calculator shows how the same token counts translate into very different monthly bills. For the full story of how words become tokens — and why 1,000 words never means 1,000 tokens — see our tokens explainer.
What a thousand tokens costs on each model
Token counts only turn into money once you attach a price to them. Here is what 1,000 input tokens and 1,000 output tokens cost across the catalogue at August 2026 list prices, per million tokens (input / output).
At the top: GPT-5.6 Sol $5.00 / $30.00 and Claude Opus 4.8 $5.00 / $25.00. The workhorse middle: Claude Sonnet 5 $3.00 / $15.00, Kimi K3 $3.00 / $15.00, GPT-5.6 Terra $2.00 / $12.00 and GLM-5.2 $1.40 / $4.40. The budget end: Gemini 2.5 Pro $1.25 / $10.00, Claude Haiku 4.5 $1.00 / $5.00, Nemotron 3 Ultra $0.50 / $2.20, Gemini 2.5 Flash $0.30 / $2.50, Llama 4 Maverick $0.20 / $0.60 and GPT-5.6 Luna $0.20 / $1.20 — Luna and Terra were cut by 80% and 20% on 30 July 2026, which is why Luna now sits at the bottom of the table rather than in the middle.
Divide by a thousand for the per-1k figure: 1,000 output tokens cost $0.030 on Sol and $0.0006 on Maverick — a 50× spread on identical token counts. That spread, not the tokenizer, is where the money is. Price your own volumes in the OpenAI and Claude calculators.
Real texts and their token counts
Abstract ratios are easy to forget, so here are shapes you probably work with. A tweet-length message of about 280 characters lands near 70 tokens. A dense paragraph of 100 English words runs roughly 130–140 tokens; the same paragraph in Portuguese usually costs 15–25% more, because accented characters and longer words split into more pieces. A one-page memo of 500 words is about 650–700 tokens.
Scaling up: a 3,000-word blog post is roughly 4,000 tokens, a 30-page report around 20,000, and a 300-page book somewhere near 200,000 tokens — which is exactly the full context window of Claude Opus 4.8, and about a fifth of GPT-5.6's 1.05M window.
Code behaves differently again: indentation, braces and identifiers fragment heavily, so 100 lines of TypeScript often costs 1,200–1,500 tokens where 100 lines of prose would cost far less. Check what fits in a given model with the context window calculator, or reason in pages with how many pages fit.
Frequently asked questions
Is 4 characters per token accurate for every language?
No — it's a reasonable approximation for English. Languages with different scripts, like Chinese, Japanese or Korean, or heavy use of special characters typically use more tokens per character. For exact counts, use the provider's official tokenizer.
Why is the same text priced differently as input vs output?
Output tokens are generated one at a time and cost more compute per token than reading input, so providers charge more for output. The same amount of text can cost several times more if the model generates it versus if you send it as a prompt.
Does formatting like markdown or code affect token count?
Yes. Structural characters, code syntax, and repeated whitespace all consume tokens just like regular text. Dense, information-rich text is more token-efficient than lightly formatted prose of the same word count.
Do images and audio also count as tokens?
Yes, in a converted form. Vision-capable models translate images into token equivalents based on resolution, and audio models bill per audio token or per minute depending on the provider. These usually have their own rates, separate from text-token pricing, so check the provider's pricing page for multimodal workloads.
Does the same text cost the same number of tokens on every model?
No. Each model family ships its own tokenizer, so identical text splits differently — the counts here are a well-calibrated estimate, not a per-vendor exact figure. Differences of 5–15% between families are normal, and larger for code, non-Latin scripts and heavily accented text. For billing-critical work, count with the provider's own tokenizer and treat this page as the planning number.
Why does Portuguese cost more tokens than English?
Tokenizers are trained on corpora that skew heavily English, so common English words often map to a single token while their Portuguese equivalents split into two or three pieces. Accents add fragmentation of their own. In practice the same content costs roughly 15–25% more tokens in Portuguese, which means the same prompt is genuinely more expensive to run — worth accounting for when you budget a bilingual product.
Are spaces and punctuation billed?
Yes. Tokenizers usually attach a leading space to the word that follows it, so whitespace is folded into the token count rather than being free. Punctuation, newlines and markdown syntax all consume tokens too. This is why heavily formatted prompts — tables, long bullet lists, JSON with generous indentation — cost more than the same information written plainly.
Curious where these numbers come from? Read our pricing methodology.