TokenAtlasWhat is a token?
A token is a chunk of text — roughly 4 characters or 0.75 words in English. AI models read and write in tokens, and providers bill per million tokens consumed.
Why output costs more
Generating a token requires running the entire model. Reading a token is comparatively cheap. That's why output is typically 3–5× input price.
How to estimate your tokens
English text: word_count / 0.75. Code: word_count × 1.5. Non-English: word_count × 2 (conservative). Or use any provider's tokenizer for an exact count.
Per-million rates make comparison easy
All major providers price per 1M tokens. To compare: (cost / 1M) × your_volume. TokenAtlas does this automatically across every model.
Worked example across three models
For a 2,000-token input and 500-token output call: GPT-4o mini ($0.15/$0.60) costs 2000/1e6*0.15 + 500/1e6*0.60 = $0.0003 + $0.0003 = $0.0006. DeepSeek V3 ($0.27/$1.10) costs $0.00054 + $0.00055 = $0.00109. Claude 3.5 Sonnet ($3/$15) costs $0.006 + $0.0075 = $0.0135 — over 20x the mini price for the identical token volume, purely from rate differences.
Mechanics: why tokenisation isn't 1 token = 1 character
Tokenizers split text into subword units learned from training data, so common English words are often a single token while rare words, numbers, and non-Latin scripts split into several. This is why the same sentence can produce different token counts across providers' tokenizers, and why code or Japanese text can cost 2-3x more per character than plain English at the same word count.
Decision checklist for comparing rate cards
— Are you comparing per-1M-token rates, not per-request or per-word figures? — Have you applied your actual input/output mix rather than assuming 50/50? — Does the model have a context-length pricing tier that changes the rate above a threshold? — Are you including both cached and uncached rates if the provider offers caching?
What this explainer does not cover
This post explains the mechanics of token-based pricing, not how to retrieve your own historical usage — that requires the provider's console. It also doesn't cover image, audio, or embedding token conversion ratios in depth beyond noting they exist; those deserve their own treatment given how provider-specific the conversion factors are.
Translating token cost into a per-output-unit figure
Once you know the cost per call, divide by whatever unit of output matters to you — per page, per report, per generated image caption — to get a figure comparable across models and against a manual-labor baseline. A $0.0135 Sonnet call producing one 500-word draft is $0.0135/draft; multiply by expected monthly draft volume to get the number that belongs in a budget.
Re-deriving the math when a provider changes rates
Providers revise per-token rates periodically, sometimes lowering costs on established models as newer ones ship. When a rate changes, keep the token-count formulas on this page fixed and substitute only the new $/1M figures — the tokenization and volume estimation steps don't change, so a rate update is a one-line substitution, not a full re-derivation of the workload.
Frequently asked questions
- Are images and audio tokenized?
- Yes — at fixed conversion ratios (e.g. a 1024×1024 image ≈ 765 tokens on GPT-4o).
- Do tools/function calls count?
- Yes — tool schemas and tool responses are billed as input and output tokens.
- Is there a cheaper alternative to tokens?
- Embedding APIs bill per token but at fractional rates (~$0.02 / 1M). Use them for retrieval, not generation.