AI API Pricing Explained: Tokens, Tiers and Why the Same Model Costs Different Amounts
- Standfirst (first paragraph, distinct from meta): "A token is the small piece of text an AI model reads or writes, and API pricing charges you separately for the tokens you send in and the tokens the model sends back.
A token is the small piece of text an AI model reads or writes, and API pricing charges you separately for the tokens you send in and the tokens the model sends back.
Key points
- Every major API prices per million tokens, split into input and output: in the table below, output costs five or six times the input rate on every example model.
- Cached input is far cheaper than fresh input — as low as $0.01 per million tokens on gpt-6-luna — and batch processing halves the bill on Anthropic’s and Google’s rates.
- A chatbot subscription does not pay an API bill: the apps and the APIs are separate products with separate pricing pages, linked only by occasional included API credits.
A token is the unit the whole bill is built on
Models do not charge per question or per word; they count the pieces of text that flow in and out.
If you have ever opened an API pricing page and felt your eyes slide off the numbers, you are not alone. Those pages are written for developers doing sums, and the first sum — what is a token, exactly? — is the one everything depends on. A token is a small chunk of text: sometimes a whole short word, sometimes part of a longer one. Your prompt is broken into tokens on the way in, and the answer is counted in tokens on the way out. Our AI terms glossary covers the rest of the vocabulary.
The pages treat that counting seriously. Google warns that on its newest Gemini models every token processed is counted and billed, including thinking tokens. Anthropic notes its newer tokenizer produces approximately 30% more tokens for the same text on Claude 4.7 and later models, so the same paragraph can cost differently across generations. Tokens are the metre reading of AI: not intuitive, but consistent.
Input and output are priced separately, and output is the expensive half
Reading your prompt is cheap; writing the answer is where the money goes.
Every rate card below splits the price in two. Input tokens are what you send: your question, documents and instructions. Output tokens are what the model generates back. On every example here, output costs five or six times the input rate, which is why a short question with a long answer can cost more than a long document with a short one. Prices per million tokens in US dollars, standard paid rates, checked 11 November 2026.
| Model (as labelled on the official page) | Input / 1M tokens | Cached input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
| OpenAI gpt-6-astra | $10.00 | $1.00 | $50.00 |
| OpenAI gpt-6.1-sol | $2.00 | $0.10 | $10.00 |
| OpenAI gpt-6-luna | $0.10 | $0.01 | $0.50 |
| Claude Sonnet 5.5 | $2.00 | $0.10 (cache hit) | $10.00 |
| Claude Opus 5.5 | $4.00 | $0.20 (cache hit) | $20.00 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.01 (cache hit) | $0.50 |
| Gemini 4 Argon (price through 30 November 2026) | $2.00 | $0.10 | $10.00 |
| Gemini 3.1 Pro Preview (prompts up to 200,000 tokens) | $2.00 | $0.20 | $12.00 |
| Gemini 3.1 Flash-Lite (text, image and video input) | $0.25 | $0.025 | $1.50 |
Cached input discounts reward sending the same context twice
When a prompt repeats, the model can re-read it from cache at a fraction of the fresh-input price.
Many real applications send the same long instructions or conversation history with every request. Prompt caching lets the provider keep that repeated portion ready, and the pricing pages pass part of the saving on. Anthropic explains it directly: caching reuses previously processed portions of your prompt, and a cache hit costs 10 percent of the standard input price on most models, 5 percent on Claude Opus 5.5 and Sonnet 5.5, and 2.5 percent on Claude Fable 5.2. Writing to the cache costs extra — 1.25 times the base input price for five minutes, twice for one hour — so caching pays off only when the context is read again. OpenAI and Google show the same principle in the table’s cached-input column — Google also adds a storage price of $1.00 per million tokens per hour on Gemini 3.1 Flash-Lite.
Batch, speed and prompt-length tiers move the same model’s price again
When your work can wait, or your prompt is unusually long, the rate card changes underneath you.
Anthropic’s Batch API offers a 50 percent discount on input and output tokens for asynchronous processing of large volumes. Google makes the same offer in plainer language: its paid tier includes Batch API at a 50 percent cost reduction, and its Gemini 3.1 Flash-Lite table shows batch input at $0.125 and batch output at $0.75, against the standard $0.25 and $1.50. Speed runs the other way: Google’s Priority tier for Flash-Lite charges $0.45 input and $2.70 output.
Prompt length can move the price too. Google charges Gemini 3.1 Pro Preview at $2.00 input and $12.00 output up to 200,000 tokens, then $4.00 and $18.00 above it. Anthropic prices Claude Haiku 5.5 the same way, rising from $0.10 and $0.50 to $0.50 and $2.50 above 100,000-token prompts. Some prices are simply dated: Gemini 4 Argon rises from $2.00 and $10.00 to $4.00 and $20.00 on 1 December 2026. A rate card is a snapshot, not a promise.
A subscription and an API account are two different products
Paying for the chatbot app does not fill your API meter, and the pricing pages keep the two apart.
This misunderstanding generates endless frustrated forum posts, so it is worth stating plainly. Subscriptions buy you the app: a chat window, usage limits and features for a fixed monthly fee. The APIs are developer products with their own pricing pages, accounts and a meter that runs per token. Our pillar guide on how AI companies make money sets the two side by side as separate revenue streams, and our explainer on AI subscriptions and pricing covers the app plans themselves.
The companies blur the line only slightly. Anthropic explains that at a paid plan’s limit you can turn on usage credits to keep working at standard API rates — the meter exists underneath, and the credit is a bridge onto it. Max includes $100 a month in API credits, Ultra $200. Google keeps Gemini subscriptions and Developer API pricing on separate pages, the API offering its own free tier — free tokens on certain models, with content used to improve Google’s products — and a paid tier where content is not used that way.
What this means
The headline price of a model is the least useful number on the page; your real cost hides in output, caching and timing.
If you are choosing an API for a project, the temptation is to compare one number — the input price — and pick the smallest. Please resist it. Most real workloads are dominated by output, because answers and generated documents are usually longer than the instruction that asked for them, and output is priced at five or six times input on every row above. A model that looks twice as expensive on input can be cheaper overall if it answers in half the words, or dearer if it thinks out loud at length, thinking tokens included.
The second lesson is kinder: the discounts are real, and they reward ordinary good habits. Cache the instructions instead of rebuilding a giant prompt for every call. Let the overnight report run on batch rates instead of paying for instant answers nobody will read until morning. Save the flagship for the questions that need it, and let a small, fast model handle sorting, tagging and first drafts. None of that requires cleverness — just reading the rate card the way you would read a phone contract, slowly, suspecting the interesting lines are the small ones.
In our view, this is also why the same model seems to cost different amounts wherever you meet it. Inside a subscription, its cost hides inside a monthly fee and a usage limit. Inside an app someone else built, it is wrapped into that app’s price. Inside the API, it is naked: per token, per direction, per tier, changing with your prompt length and your patience. That looks unfriendly, but it is honest, and once you can read it you can predict a bill before it arrives — the difference, for anyone building on these models, between an experiment and a business. Check the date before you budget, though: as Gemini 4 Argon’s December rise shows, November’s number may not be December’s.
Bottom line
Read any API price as four numbers — input, output, cached input and tier — and the mystery mostly disappears.
Tokens are the metre, output is the expensive direction, caching and batch work are the honest discounts, and your subscription was never going to pay this bill. Compare models on your actual workload, and re-check the official page before the project, not after the invoice.
Related Guides
Start with the pillar, then brush up on the vocabulary and subscription pricing.
- How AI Companies Make Money: Subscriptions, APIs and Enterprise Deals Explained
- AI Terms Glossary: Tokens, Prompts, RAG, Hallucinations and More, Explained
- AI Subscriptions and Pricing Explained
Sources
The official API pricing pages behind the table and every rate above.
