MTok means a million tokens, the unit every AI API price is quoted in. How tokens, context windows, batch processing and prompt caching change what you pay.
A token is the unit AI models use to process text — roughly 4 characters, or about three-quarters of an English word. APIs bill per token, so a 1,000-word prompt is around 1,300 input tokens. Every price on this site is quoted per million tokens (MTok).
MTok is one million tokens, the standard unit for AI API pricing. If a model costs $2.50/MTok input, then one million tokens of input — roughly 750,000 words — costs $2.50. Input (your prompt) and output (the model's response) are billed at different rates.
Generating text is far more compute-intensive than reading it, so most providers price output 3–6x above input — Claude Opus 5.5, for example, charges $4.00/MTok in but $20.00/MTok out. If your workload produces long responses, output pricing will dominate your bill.
The context window is the maximum amount of text (in tokens) a model can consider at once — your prompt, any documents, and the conversation so far. Current flagships like Claude Fable 5.2 and Gemini 4 Carbon support 1M tokens, and OpenAI's GPT-6 Astra and GPT-5.x models reach 1.05M; Meta's Llama 4 Scout has a 10M native window, though hosts often serve less. Bigger windows let you process whole codebases or document sets in one call.
Batch processing queues API requests for asynchronous completion (typically within 24 hours) instead of answering in real time. Because providers can schedule the work on spare capacity, Anthropic, OpenAI, Google, and Mistral discount batch jobs by 50% on both input and output tokens, and Azure does the same for most GPT models (though not yet for GPT-6 Astra). The rate is not universal: xAI discounts by only 20%, and then only on Grok 4.7, Grok 4.6, Grok 4.5, Grok 4.3 and the Grok 4.20 variants, while DeepSeek has no batch API at all, though its off-peak hours bill at half price. Anything that doesn't need an instant response — ETL, evals, backfills — belongs in a batch.
Prompt caching lets you reuse a repeated prompt prefix (system prompts, shared documents) at a fraction of the normal input rate. OpenAI bills cached GPT-6 Astra input at $1.00/MTok versus $10.00 list, Anthropic's Claude Fable 5.2 cuts cache reads to 2.5% of input ($0.20/MTok) and Opus 5.5 and Sonnet 5.5 to 5% ($0.20 and $0.10/MTok), and DeepSeek V4.1 Flash cache hits cost under a cent per MTok. Check whether your model charges to write the cache as well as read it: OpenAI's GPT-6 and GPT-5.6 models bill cache writes at 1.25x the input rate, so prompts that are refreshed constantly rather than reused can cost more, not less. If many requests share the same preamble, caching is often the single biggest saving available.
Rules of thumb for English: 1 token ≈ 4 characters, 1,000 tokens ≈ 750 words, and a typical A4 page is 500–800 tokens. Code and non-English text usually tokenize less efficiently. Tokenizers also differ per model family — Anthropic's current tokenizer, for instance, can produce noticeably more tokens for the same text than older Claude versions.
Consumer subscriptions (ChatGPT Plus, Claude Pro) are flat monthly fees for using a chat app, with usage limits. API pricing is pay-per-token for building the models into your own products and workflows — there's no monthly fee, you pay exactly for what you process. Manus is a third pattern: credit-based plans where tasks consume credits.