AI Tokens Explained: Why Prices Are Per Million, Not Per Word

AI tokens explained, starting with the number that makes it concrete: Claude Opus 5, released July 24, 2026, costs $5 per million input tokens and $25 per million output tokens. A token is not a word. It’s a chunk of text, roughly four characters or about three-quarters of a word in English. So a million tokens is somewhere around 750,000 English words, or ten decent novels.

That ratio, 1 to 5 between reading and writing, is the single most useful fact a family can know about AI pricing. It means asking a model to read a long document is cheap. Asking it to produce a long document is expensive. Most people’s intuition has it backwards.

Key Takeaways

• A token is a text chunk produced by a tokenizer, roughly 4 characters or 0.75 English words. Models process tokens, not words or letters.
• Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. Fable 5.1 is $10 in and $50 out. Haiku 4.5 is $1 in and $5 out.
• Output is consistently about five times the price of input across tiers, because generating text requires a forward pass per token while reading it can be processed in parallel.
• Tokenizers differ between models. Anthropic’s docs note that Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text, which changes the per-token math even at identical prices.
• The reason AI struggles to count letters in a word is tokenization: it sees chunks, not characters. This is a great teaching moment and a genuine limitation.

What a token actually is

A tokenizer is a program that chops text into pieces a model can handle. It runs before the model sees anything. The pieces it produces are tokens, and the model’s entire world is a sequence of them.

Why not just use words? Two reasons. First, vocabulary size. English has hundreds of thousands of words, plus names, typos, and other languages. A model would need an entry for each. Second, generalization. If “running” and “runs” are unrelated entries, the model has to learn each independently.

The fix, subword tokenization, comes from Sennrich, Haddow, and Birch’s 2016 byte-pair encoding paper, and it splits text into frequent chunks. Common words become one token. Rare words split into pieces. “Cat” is one token. “Tokenization” might be three. A name like “Xochitl” might be four or five. Punctuation and spaces count.

The Anthropic pricing documentation gives the working estimate: “1 token is approximately 4 characters or 0.75 words in English. The exact count varies by language and content type.”

That last clause matters more than it looks. Tokenizers are trained on data that skews English, so Spanish, Portuguese, and especially languages with non-Latin scripts often use more tokens for the same meaning. Same content, higher bill.

And here’s the fact that explains a behavior your kid has definitely noticed: models are bad at counting the letters in a word, or at telling you how many r’s are in “strawberry,” because they don’t see letters. They see chunks. Asking a model to count characters is like asking someone reading a book through a stack of index cards to count individual pen strokes.

Why output costs five times input

Look at the actual published price table and the pattern is unmistakable:

Opus 5: $5 in, $25 out. Fable 5.1: $10 in, $50 out. Sonnet 5: $2 in, $10 out. Haiku 4.5: $1 in, $5 out. Every tier holds the 1-to-5 ratio.

The mechanical reason is how transformers work. When you send input, the model can process the whole prompt in parallel; all those tokens go through the network together in one pass. When the model generates output, it produces one token at a time, and each new token requires another pass through the entire network, because it depends on everything generated so far. Generating 1,000 tokens means 1,000 sequential passes. Reading 1,000 tokens means roughly one.

The consequences for a family using AI:

Reading is cheap. Pasting a 20-page PDF and asking a question costs very little. Roughly 10,000 tokens of input at Opus 5 rates is five cents.

Writing is expensive. Asking for a 3,000-word essay generates roughly 4,000 output tokens, which at $25 per million is ten cents. Cheaper tiers change the arithmetic sharply: Google’s August 2026 roundup put Gemini 3.7 Flash at half the prior tier’s per-token price. Sounds small until it’s a class of 30 doing it daily.

Thinking is output. This is the one people miss. A reasoning model’s internal deliberation is generated tokens, billed as output. A model that thinks for 40 minutes on a hard problem can produce tens of thousands of output tokens before writing its answer. That’s the cost side of test-time compute.

Conversation history compounds. Every turn resends the whole prior conversation as input. Turn 20 of a chat sends everything from turns 1 through 19 again. It’s cheap per turn, and it adds up, which is exactly why prompt caching exists; the caching documentation spells out the multipliers.

The tokenizer change nobody mentions

Here’s a detail buried in the Anthropic docs that has real financial consequences: “Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer that contributes to their improved performance on a wide range of tasks. This tokenizer produces approximately 30% more tokens for the same text.”

Read that carefully. Same text, 30% more tokens. So even with identical posted prices, a newer model can cost more per page of actual content. The tokenizer change bought better performance and a higher token count for the same input.

This is a genuinely useful thing to understand, because it means “price per million tokens” is not directly comparable across models with different tokenizers. When Sam Altman said GPT-5.6 Sol was “54% more token efficient” on coding, that’s a claim about the same axis from the other direction. Our piece on how to read AI marketing numbers like token efficiency works through the sleight of hand.

The honest summary for a parent: the sticker price is one of two numbers that determine your bill, and the other one, tokens per page of your actual content, is rarely published.

How to Teach Your Kid About Tokens

Tokens are a counting problem, which makes them ideal for kids. Everything here uses paper.

Ages 5–8: Chop the sentence

Write a short sentence on paper: “The cat sat on the mat.” Cut it into words with scissors. Count the pieces: six. Now write a longer word, like “unbelievable,” and cut it into three chunks: “un,” “believ,” “able.” Say: “Computers read like this. They cut everything into little pieces and count them. Big words become more pieces.” Then ask them how many pieces are in their own name. They’ll cut it into two or three, which is exactly what a tokenizer does with unusual names.

Ages 9–12: The four-character rule

Give your kid a paragraph from a book, roughly 100 words. Have them count the characters including spaces (or measure it: about 5.5 characters per word average, so estimate). Divide by 4. That’s the approximate token count, and it should land near 130 to 140 for a 100-word paragraph. Now do the math: at $25 per million output tokens, what would it cost for the AI to write that paragraph? (About a third of a cent.) Then: how many paragraphs before it costs a dollar? Around 300. Kids who do this division once never think of AI as free again.

Ages 13+: Measure the input-output ratio

Have your teen run a real experiment. Take one long article. Task A: ask an AI to summarize it in three sentences. Task B: ask the AI to write a 1,000-word essay on the same topic. Estimate tokens for both (input and output separately, using the 4-character rule) and compute the cost at Opus 5 rates. They’ll find Task A is dominated by input and costs little; Task B is dominated by output and costs more despite having a shorter prompt. Then have them find the tokenizer note in the Anthropic pricing page and explain why two models at the same posted price can cost different amounts.

The question to ask: “Is this task mostly reading or mostly writing? Which one costs more?”

Text to tokens to dollars: a worked table

Text	Approx. characters	Approx. tokens	As input at Opus 5 ($5/M)	As output at Opus 5 ($25/M)
“cat”	3	1	$0.000005	$0.000025
One tweet-length sentence (~140 chars)	140	~35	$0.00018	$0.00088
One paragraph (~100 words)	~550	~140	$0.0007	$0.0035
A 5-page essay (~2,500 words)	~13,750	~3,400	$0.017	$0.086
A 300-page novel (~90,000 words)	~495,000	~124,000	$0.62	$3.10
One million tokens	~4,000,000	1,000,000	$5.00	$25.00

Rates from Anthropic’s published pricing for Claude Opus 5. Token counts use the documented 4-characters-per-token estimate and will vary by content and tokenizer.

Notice the novel row. Handing a model an entire 300-page book to read costs about 62 cents. Asking it to write one costs over three dollars, plus it would take hours. The asymmetry is the lesson.

What to do at home

Ask for shorter outputs on purpose

“Answer in three sentences” is not just good for comprehension. It’s a direct cost control, since output is five times the price of input. Teaching a kid to specify length is teaching them to be a better user and a cheaper one.

Paste generously, ask concisely

If your kid is working from a source document, paste the whole thing. Input is cheap and grounding an answer in the actual text reduces errors. Then ask a narrow question. This is the opposite of what most people do.

Watch out for thinking modes on long tasks

A high-effort reasoning setting generates a large number of output tokens before answering. On a paid plan that’s real money. Use it when the problem needs it and turn it down for routine work.

Make the division once

Sit down with your kid and calculate what your family’s AI usage actually costs in tokens for one week. It takes fifteen minutes and it permanently changes how both of you think about “just asking the AI.” Most families find the number surprisingly small, which is also useful information.

What not to do

Don’t tell your kid a token is a word. It’s close enough for estimating and wrong in the way that matters: it’s why the model can’t count letters, why non-English text costs more, and why two models at the same price can bill differently. The chunk concept is barely harder and it explains actual behavior.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can estimate the token count of a paragraph using the 4-character rule, and knows output costs more than input.
• Month 2 red flags: They ask for maximum-length answers by default. Or they’re confused when the model miscounts letters in a word and treat it as the model being broken.
• Month 3 self-check: Hand them two tasks and ask which will cost more. If they reason from input-versus-output rather than from how hard the task sounds, they’ve got it.

Frequently Asked Questions

How many words are in a million tokens?

Roughly 750,000 English words, using Anthropic’s documented estimate of about 0.75 words per token. That’s approximately ten novels. The count varies by language and content type, and non-English text typically uses more tokens for the same meaning.

Why is output more expensive than input?

Because generation is sequential. Each output token requires a full pass through the network and depends on every token before it, while an input prompt can be processed in parallel in roughly one pass. Across Claude’s tiers, output is priced at five times input.

Why can’t AI count the letters in “strawberry”?

Because it never sees letters. The tokenizer splits text into chunks before the model reads it, so the model processes something like “straw” plus “berry” rather than eleven characters. Character-level tasks are genuinely awkward for token-based models.

Does the same text always produce the same number of tokens?

No. Different models use different tokenizers. Anthropic’s docs note that Claude 4.7 and later use a newer tokenizer producing roughly 30% more tokens for the same text than earlier models. So identical posted prices can still mean different real costs.

Should I worry about token costs for my kid’s homework help?

Usually not, at typical household volume. A five-page essay generated at Opus 5 rates costs under ten cents. The costs that surprise people are long reasoning sessions, agent tasks that run for a while, and consumer subscriptions where you’re paying a flat fee regardless. Understanding tokens tells you which situation you’re in.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Anthropic. (2026). “Pricing.” Claude Platform Documentation. https://platform.claude.com/docs/en/about-claude/pricing
2. Gizmodo. (2026, July 24). “Anthropic releases new Claude model, positions it as a cost-efficient version of Fable 5.” https://gizmodo.com/anthropic-releases-new-claude-model-positions-it-as-a-cost-efficient-version-of-fable-5-2000790486
3. Sennrich, R., Haddow, B., & Birch, A. (2016). “Neural Machine Translation of Rare Words with Subword Units.” Proceedings of ACL 2016. https://arxiv.org/abs/1508.07909
4. Anthropic. (2026). “Prompt caching.” Claude Platform Documentation. https://platform.claude.com/docs/en/build-with-claude/prompt-caching
5. OpenAI. (2026, July 9). “GPT-5.6.” https://openai.com/index/gpt-5-6/
6. Google. (2026, August). “Google AI updates, August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
7. MarkTechPost. (2026, September 1). “Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1.” https://www.marktechpost.com/2026/09/01/anthropic-releases-claude-fable-5-1-and-claude-mythos-5-1-52-6-on-terminal-bench-science-and-75-cheaper-cache-reads/