Cost per token is the price an AI model vendor charges for each token of text it reads or writes. It's quoted per million tokens, with separate input and output rates. Cost per task = (input tokens × input price + output tokens × output price) ÷ 1,000,000. As at October 2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens.
Key takeaways
- Prices are quoted per million tokens (MTok), separately for input and output.
- Cost per task = (input tokens × input price + output tokens × output price) ÷ 1,000,000.
- Output tokens cost about five times input at OpenAI, Anthropic and Google.
- Caching and batch discounts can cut a bill by half or more, so always model them.
Two cents.
That's roughly what it costs Claude Sonnet 5.5 to read a 2,000-word brief and write a 1,000-word blog post from it. A human writer would charge you a lot more than two cents. (I certainly would.)
That gap, between what a task costs you and what it's worth to a customer, is the whole game in AI pricing. And it starts with one number: cost per token.
What does cost per token mean?
Cost per token is what an AI model vendor charges for each token of text the model reads (input) or writes (output). A token is about four characters, or three-quarters of a word.
Because a single token costs a tiny fraction of a cent, vendors quote prices per million tokens, written "per MTok."
The cost per token formula
Cost per task = (Input tokens × Input price per 1M + Output tokens × Output price per 1M) ÷ 1,000,000
Multiply by tasks per month to get your monthly bill.
What does a 1,000-word blog post cost?
The brief is about 2,000 words (roughly 2,700 input tokens). The post is 1,000 words (about 1,350 output tokens).
On Claude Sonnet 5.5 ($2 in, $10 out): 2,700 × $2 ÷ 1M = $0.0054, plus 1,350 × $10 ÷ 1M = $0.0135. About $0.019 a post.
On Claude Opus 5.5 ($4 in, $20 out): about $0.038 a post.
On GPT-6 Luna ($0.10 in, $0.50 out): about $0.001 a post, if the small model is good enough.
Notice that output is two-thirds of the cost even though it's half the tokens.
Cost per token in October 2026
| Model | Input per 1M | Output per 1M | Per single output token |
|---|---|---|---|
| GPT-6 Astra / Claude Fable/Mythos 5.1 (frontier) | $10 | $50 | $0.00005 |
| Claude Opus 5.5 | $4 | $20 | $0.00002 |
| Claude Sonnet 5.5 / GPT-6.1 Sol | $2 | $10 | $0.00001 |
| GPT-6.1 Terra | $1 | $5 | $0.000005 |
| Claude Haiku 5.5 / GPT-6 Luna | $0.10 | $0.50 | $0.0000005 |
Prices checked 27 October 2026. Gemini 4 Argon is scheduled to rise from $2/$10 to $4/$20 on 1 December 2026, so cost your product at the higher price if you build on it. For every model, see the LLM API pricing comparison or the live AI Pricing Index.
AI Token Cost CalculatorEstimate monthly LLM API spend across models.Open the free tool →What changes the real cost per token?
- Caching. Anthropic charges cache hits at about 5% of the input price. OpenAI discounts cached input 90–95%. Repeated instructions get very cheap.
- Batch processing. Jobs that can wait up to 24 hours are 50% off at OpenAI, Anthropic and Google.
- Surcharges. Priority processing, very long prompts and data residency can add 10% or more.
- Hidden tokens. Conversation history is resent on every call, and reasoning models bill their thinking as output.
Common confusions
Cost per token vs price per token. Cost is what you pay the vendor. Price is what you charge your customer. If those are the same number, you're a reseller with no margin.
Per 1,000 vs per million. $0.002 per 1,000 tokens is $2 per million. Check the unit before you compare.
Token cost vs total cost to serve. Hosting, data APIs, payment fees and support all sit on top. Your AI gross margin has to cover the lot.
Why this matters for your margins
In Chapter Five of The Subscription Playbook, high gross margin is one of the "Toughness" factors in the MOATS framework. Traditional software ran at 75–80% gross margin because serving one more customer cost almost nothing. AI changed that. Every click now has a cost per token behind it.
Work out your cost per task
The free AI Token Cost Calculator turns your prompts and volumes into a monthly bill across models.
What to do next
Pick your three most common tasks, count their tokens, and run the formula on two or three models. Then model your margins with the AI Gross Margin Calculator.
For caching, routing and whether to pass token costs on to customers, read token-based pricing explained.
Back to the subscription and pricing glossary.
Frequently asked questions
What is a blended cost per token?
A single average price that combines input and output at your actual mix. If your tasks send three input tokens for every output token on a $2/$10 model, the blended rate is (3 × $2 + 1 × $10) ÷ 4 = $4 per million tokens. It's handy for quick forecasts, but recalculate it whenever your prompt or answer lengths change.
Why are some prices quoted per 1,000 tokens?
Older price lists and some cloud platforms quote per 1,000 tokens. To convert, multiply a per-1,000 price by 1,000 to get the per-million price. $0.002 per 1,000 tokens is the same as $2 per million. Always check which unit a quote uses before you compare two vendors.
Are token prices the same in Australian dollars?
Most model vendors bill in US dollars, so an Australian business's cost per token moves with the exchange rate. If you charge customers in A$, build a buffer into your prices and review them every quarter. GST treatment of imported digital services also depends on your registration, so check with your accountant.
Does cost per token include hosting and other costs?
No. It's only the model vendor's charge. Your real cost to serve a customer also includes hosting, vector databases, search or data APIs, payment fees and support. Add those on top when you work out your AI gross margin, or your margins will look better on paper than in your bank account.
Get both books free
The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.
Robert Coorey, MBA
Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.
Sources
- OpenAI: API pricing (checked 27 Oct 2026)
- Anthropic: Claude API pricing (checked 27 Oct 2026)
- Google: Gemini Developer API pricing (checked 27 Oct 2026)
Facts and prices checked 27 October 2026. Spotted something out of date? Tell us.