How Does AI API Pricing Work? Token Pricing Explained in Plain English

Photo: Hippopx (public domain). Illustrative image.
What is a token?
A token is the unit AI models count in. Roughly, one token is about 4 characters of English text, or about three-quarters of a word — so 1,000 tokens is about 750 words. A short question might be 20 tokens; a long document pasted into a chat might be 10,000. The model does not charge per question, per minute or per call. It charges per token, in and out, and the bill is the sum of the two. The formula is simple: cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price). Everything else in this guide is about the three numbers in that formula: how many tokens you use, and what each side costs.
Why are input and output priced separately?
Because they cost the provider different amounts of work. Input tokens are processed in parallel — the model reads your whole prompt at once. Output tokens are generated one at a time, each one depending on the one before it, which keeps the expensive hardware busy for longer. That is why output is almost always dearer: typically 3–8 times the input price on flagship models, and 5–6 times on the most expensive ones. On OpenAI’s pricing page, as listed on 17 November 2026, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens; GPT-6 Luna, the budget sibling, costs $0.10 and $0.50. On Anthropic’s page the same day, Claude Sonnet 5.5 costs $2 and $10, and Claude Fable 5.2 costs $8 and $40. The ratio is the pattern, not the exact number — output is where the money goes.
How much does a typical request cost?
Take a 10,000-token prompt that gets a 2,000-token reply. On GPT-6 Luna ($0.10 input, $0.50 output per million tokens), the input costs 10,000 ÷ 1,000,000 × $0.10 = $0.001, and the output costs 2,000 ÷ 1,000,000 × $0.50 = $0.001 — two-tenths of a cent in total. The same request on GPT-6 Astra ($10 and $50) costs $0.10 + $0.10 = $0.20. On Claude Sonnet 5.5 ($2 and $10) it costs $0.02 + $0.02 = $0.04. The lesson is not that one model is “cheap” and another “expensive” in the abstract — it is that the model choice is the price. The cheapest model in the range, Tencent’s Hy4 Flash at $0.07 per million input tokens, would do the same request for about a tenth of a cent. Pick the model for the job, then read its price — not the other way round.
What is prompt caching, and how much does it save?
Prompt caching is a discount for repetition. If you send the same long prompt many times — a system instruction, a document, a set of examples — the provider stores the processed version and charges you far less for reusing it. On OpenAI’s page, cached input for GPT-6 Astra costs $1.00 per million tokens against $10.00 uncached — a tenth of the price. On Anthropic’s page, a cached read for Sonnet 5.5 costs $0.10 against $2.00 input. There is a catch: writing to the cache costs extra (Anthropic charges a cache-write fee, and OpenAI charges more for longer cache lifetimes), and cached entries expire — Anthropic’s pricing reflects a 5-minute time-to-live. Caching pays when the same prompt is reused many times; it does nothing for a one-off question.
What is batch processing, and when should I use it?
Batch processing is a discount for patience. If your work does not need an answer now — overnight evaluation, bulk document processing, a queue of jobs — you can submit it as a batch and pay about half price. OpenAI’s Flex/Batch pricing for the GPT-6 and GPT-6.1 models is exactly half the standard input, output and cached-input prices (GPT-6 Astra at $5.00 and $25.00 instead of $10.00 and $50.00), and Anthropic advertises “save 50% with batch processing”. The trade-off is time: batch jobs run when capacity is available, not immediately. Use it for anything that can wait a few hours; never use it for anything a user is waiting on.
Why do prices vary so much between models?
Because the price follows the model’s cost to run. A small model that activates 12 billion parameters per token, like Hy4 Flash, is cheap to serve; a flagship that activates far more is not. As listed on 17 November 2026, flagship input prices ranged from $0.10 per million tokens (GPT-6 Luna, Claude Haiku 5.5 on short prompts) to $15 (GPT-5.5 Pro and GPT-5.4 Pro). Haiku 5.5 is tiered by prompt size: $0.10 per million input tokens for prompts up to 100,000 tokens, rising to $0.50 above that. There are also speed options — Anthropic offers a fast mode for Opus 5.5 and Fable 5.2 at twice the standard price for up to 2.5 times faster speeds — and regional options, such as US-only inference at 1.1 times the price. The spread is the point: there is no single “AI price”, only a price for a specific model, on a specific tier, at a specific speed.
Do reasoning models cost more?
Yes, and in a way that surprises people. A reasoning model thinks before it answers, generating intermediate “reasoning tokens” — and those tokens are billed as output, at the output price. A long chain of thought can cost more than the answer itself. That is why the cheapest way to use a reasoning model is to ask it a question it actually needs to think about, and why some providers now show the reasoning-token count separately. The same logic applies to agents: an agent that makes ten tool calls and reads ten pages is paying for every token in every step, which is why Qwen4-Max’s claim of a 62% cut in cost per task matters more than its per-token price.
How do I estimate a bill before I build?
Three steps. First, measure your tokens: send a few real requests and read the usage figures the API returns — do not guess from word counts. Second, split them into input and output, and into cached and uncached, because the prices differ by an order of magnitude. Third, apply the formula above with the current prices from the provider’s own pricing page, and add a margin for retries and agent loops, which are the usual surprise. Rates move monthly — GPT-5.6 Sol’s promotional pricing, for example, is listed as available only until 30 November 2026 — so never budget from a price quoted in an article, including this one. The figures here are as listed on 17 November 2026; the method is the part that stays true.
Frequently asked questions
What is a token, in one sentence? About 4 characters of English, or about three-quarters of a word; 1,000 tokens is roughly 750 words.
Why is output more expensive than input? Output is generated one token at a time, which ties up hardware longer; input is processed in parallel.
What is the cheapest way to run a model? The cheapest model that does the job, with prompt caching for repeated prompts and batch processing for work that can wait.
Are free tiers really free? Usually, with limits — a monthly credit allowance, non-commercial use, or restricted downloads. Read the plan page, not the headline.
How do I know what I actually spent? Every API response returns token usage; the provider’s dashboard turns that into a bill. Check it weekly while you are learning, not monthly.
Sources
- OpenAI — API pricing (developers.openai.com/api/docs/pricing, read 17 November 2026; flagship models, short context, standard speed): GPT-6 Astra $10/$50 per million input/output tokens, GPT-6 Luna $0.10/$0.50, GPT-6.1 Sol $2/$10, GPT-5.5 $2.50/$15; Flex/Batch half price for GPT-6/6.1 models
- Anthropic — API pricing (claude.com/pricing, read 17 November 2026): Fable 5.2 $8/$40, Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Haiku 5.5 $0.10/$0.50 (prompts up to 100K tokens); batch processing saves 50%
- FreshSynapse — “Tencent Hy4 Flash Matches Claude Haiku 5.5 at 1/18th the Cost Per Task”: https://freshsynapse.com/articles/tencent-hy4-flash-cheapest-ai-model-2026-11-17/
- FreshSynapse — “Qwen4-Max Is Here: Alibaba’s 2.4T-Parameter Flagship Cuts Agent Costs 62%”: https://freshsynapse.com/articles/qwen4-max-alibaba-flagship-ai-teammate-2026-11-16/
- FreshSynapse — “What Are Large Language Models?”: https://freshsynapse.com/articles/what-are-large-language-models/