AI Tokens Explained: What You're Actually Paying For

Whether you pay for a chat subscription, an API key, or an orchestration plan, the meter under the hood counts the same unit: tokens. AI tokens are the atoms of every large language model — models read tokens, think in tokens, emit tokens, and every price you will ever see is some multiple of them. Understanding tokens is the difference between buying AI deliberately and buying it blind.

What a token actually is

Models don’t read letters or whole words. Before your prompt reaches the model, a tokenizerchops it into chunks drawn from a fixed vocabulary — common words become one token (“the”, “hello”), rarer words split into pieces (“orchestration” might be “orchestr” + “ation”), and punctuation, spaces, code symbols, and emoji all count. The useful rules of thumb for English:

• 1 token ≈ 4 characters ≈ ¾ of a word
• 1,000 tokens ≈ 750 words — a long email or a short blog post
• A 300-page book ≈ roughly 100k tokens

Two caveats worth knowing. Code is token-dense: brackets, indentation, and symbols tokenize less efficiently than prose. And non-English languages often cost noticeably more tokens for the same meaning, because tokenizer vocabularies are English-heavy.

Input vs output: the two meters

Every AI call runs two meters. Input tokens are everything the model reads: your question, the system instructions, any documents you attached, and — this surprises people — the entire conversation so far, which is re-sent with every message. Output tokens are what the model writes back.

Output is priced several times higher than input across essentially every provider, because generating text is sequential, compute-hungry work while reading it is parallelizable. Two practical consequences follow directly:

• Long chats get quietly expensive. Message thirty re-sends twenty-nine messages of history as input. Starting a fresh conversation for a fresh topic is genuine cost hygiene, not superstition.
• Verbose answers cost more than verbose questions. Asking for “a table, no preamble” is a real optimization on the expensive meter.

A third meter exists on reasoning models: thinking tokens. Models that deliberate before answering bill that hidden deliberation as output, so a one-line answer to a hard question can legitimately cost thousands of tokens. Good routing sends easy questions to models that don’t overthink them.

What everyday tasks roughly cost

The shapes matter more than the exact numbers: once you see that summarization is input-heavy while code generation is output-heavy, you can predict where the money goes — and which meter to optimize.

Why the same tokens cost wildly different amounts

The per-token price spread between models is enormous — commonly ten to fifty times between a frontier flagship and a capable open-weight model like the DeepSeek, Kimi, GLM, or Qwen families. The flagship is worth it on the hardest 10–20% of your work. Paying flagship rates to reformat a table or summarize notes is the single most common way AI budgets evaporate.

This is why routing beats discipline. You will not remember to model-switch when busy; a router does it on every request. Our guide on using multiple AI models together covers the manual version of this habit, and LLM routing explained covers how the automated version decides.

Tokens vs context windows — related, not identical

The context window is the model’s working memory, measured in tokens: how much it can consider at once. Your token budget is your spend: how much reading and writing you have paid for. A model with a million-token window will happily let one giant prompt consume a meaningful slice of a monthly allowance — window size is capacity, not a discount. The full story is in our context-windows explainer.

How to make a token budget go further

And on the buying side: the three shapes of AI pricing are subscription caps (flat fee, opaque limits), raw per-token APIs (transparent, but unbounded bills), and flat token allowances (a visible monthly budget spent across many models). We compared them in one AI subscription vs many — the short version is that an explicit allowance is the most honest shape, because you can see the whole budget instead of discovering a cap mid-conversation. KernelFold’s plans work exactly this way; current allowances are on the pricing section.

The bottom line

Tokens are not a gimmick — they are the honest unit of AI work. Once you know that output costs more than input, that history is re-billed every turn, that thinking models bill their deliberation, and that the price spread between models is a full order of magnitude, most “why is my AI bill weird?” mysteries dissolve. Spend the flagship tokens where they earn their price, and let cheaper models — or a router — handle everything else.