The AI Token Glossary: Every Token Term Defined
Plain-English definitions for tokens, tokenizers, context windows, caching, and every billing term on an LLM invoice.
Updated Aug 10, 2026 · Reviewed by PromptCostLab
Token and pricing information changes frequently. Use this guide for planning and verify model-specific limits and prices with the provider before production billing.
Token
The atom of LLM computation: a piece of text treated as one unit. In English, a token is often a common word (the, and) or a word fragment (understand → under + stand, or similar splits). Every price, context limit, and usage statistic in this industry is counted in tokens. For English prose, one token ≈ 0.75 words ≈ 4 characters.
Tokenizer
The algorithm that converts text to tokens and back. Each model family trains its own tokenizer on its own data, which is why the same sentence yields different counts on GPT, Claude, Gemini, Grok, and Kimi. OpenAI publishes tokenizers (cl100k, o200k); Anthropic and Google expose counting endpoints instead. When a tokenizer generation changes — as in Anthropic's 2026 refresh — token counts for identical text can jump 30%.
Context window
The maximum number of tokens a model can process in one request, covering instructions, conversation history, retrieved documents, tool results, and the response space. Mainstream 2026 windows run 200K to ~1.05M tokens. It is working memory, not storage: the window resets between independent requests and is shared by everything in one call.
Input, output, and cached tokens
Input tokens are everything in the request you send — including system prompts, history, and tool schemas, not just the visible message. Output tokens are what the model generates, usually the most expensive category per token. Cached input tokens are repeated prompt prefixes a provider has processed recently, billed at roughly 10% of the input price (DeepSeek's hits are ~1/120th).
- Input: what you send. Billed per million, cheapest tier.
- Output: what the model writes. Typically 3–6x the input rate.
- Cached read: reused prefix at a steep discount.
- Cache write: some providers charge 1.25x input (2x for Anthropic's 1-hour cache) to store a prefix.
- Reasoning/thinking tokens: internal chain-of-thought, billed as output.
BPE (byte-pair encoding)
The training method behind most modern vocabularies: start from bytes, repeatedly merge the most frequent pairs into new tokens. Frequent patterns become single cheap tokens; rare strings stay fragmented. That is the mechanism behind every ratio in token math — why 'the' is one token and a URL is ten.
Max output tokens
A separate cap from the context window on how much the model may generate in one response: 128K on GPT-5.6 and current Claude flagships, 65,536 on Gemini, 64K on Haiku 4.5, up to 384–390K on DeepSeek V4. Long-synthesis tasks need output headroom reserved before loading evidence into the prompt.
Long-context tier
A pricing surcharge some providers apply above an input threshold: OpenAI doubles input (1.5x output) above 272K tokens on GPT-5.6; xAI doubles everything at 200K+; Gemini 3.1 Pro charges $4/$18 above 200K instead of $2/$12. A cost model that ignores tiers can understate large-prompt bills by 100%.
Prompt caching
A provider feature that stores a processed prompt prefix and reuses it on matching later requests. Reads are cheap (10–15% of input at OpenAI, Anthropic, xAI, Kimi; 90% off at Mistral; ~1/120th at DeepSeek). Google also charges hourly storage for explicit caches. The design rule: stable content first, variable content last — any change near the top invalidates everything after it.
Usage fields
The per-request accounting a provider returns: input tokens, cached tokens (read and creation), output tokens, and reasoning tokens where applicable. These are the billing ground truth. Every estimate from every calculator — including ours — should be calibrated against them in production.
Cost per million tokens
The industry's standard price unit, written $X/$Y for input/output. Examples verified August 2026: Claude Sonnet 5 $2/$10, GPT-5.6 Terra $2/$12, Gemini 3.6 Flash $0.75/$3.75 (intro), DeepSeek V4 Flash $0.14/$0.28. Cost for a request = tokens ÷ 1,000,000 × price, summed per category.
What is the definition of a token in AI?
A token is a piece of text — commonly a word fragment, whole word, or symbol — that a language model reads and generates as a single unit. Models never see raw letters; a tokenizer converts your text into token IDs first, and nearly all LLM pricing and context limits are denominated in tokens.
What is tokenization?
Tokenization is the process of splitting text into those pieces using a learned vocabulary. Byte-pair-encoding (BPE) and its variants merge frequent character sequences into single tokens, which is why common words are cheap and rare strings are expensive.
What does 'tokens' mean in an AI model's pricing?
Providers charge per million tokens, separately for input (text you send), output (text the model generates), and sometimes cached input (repeated prefixes served from cache at a discount). A model listed at $2/$10 costs $2 per million input tokens and $10 per million output tokens.
Prices, token behavior, and model limits change. These primary provider docs are the right place to verify the details before shipping.