What Are AI Tokens? The September 2026 Cost Guide to Claude and GPT

Every AI bill you will ever get from Anthropic or OpenAI is a count of tokens multiplied by a price per million. If you run AI anywhere in your business, and if you use an AI app for your Shopify store you already do, understanding that one sentence is the difference between a predictable line item and a surprise.

This guide does three things. It explains what a token is in plain terms. It lays out what the current models cost as of September 4, 2026, taken from both vendors' official pricing pages on that date. And it shows where the money actually goes in a real request, using numbers we measured on our own product-description and collection-page prompts this week, then walks through the nine levers that cut the bill.

What a token actually is

A language model does not read letters or words. It reads tokens, which are chunks of text somewhere between a character and a word. Anthropic's own rule of thumb is that one token is about four characters, or roughly 0.75 words, in English. A 300-word product description is about 400 tokens. A 2,000-word collection page is about 2,700.

Three things make tokens more slippery than that rule suggests.

Tokenizers differ by model. Anthropic states that Claude 4.7 and later use a newer tokenizer that produces approximately 30% more tokens for the same text than earlier Claude models. The same paragraph is a different token count on different models, so a token is not a stable unit across vendors. Dollars are. Compare bills, not token counts.

Images are tokens too. When you send a product photo for the model to look at, it is converted into input tokens, and how many depends on the model and the detail level you ask for. We will show a measured example below where this single fact roughly tripled the input side of a bill.

Reasoning is output. Models that think before they answer produce reasoning tokens, and both vendors bill those at the output rate even though you never see them. On one of our runs below, 805 of 1,746 output tokens were reasoning. Nearly half the output bill was the model talking to itself.

The price sheet, September 2026

All prices are per million tokens, captured from the official pricing pages on September 4, 2026. Input is what you send. Output is what comes back, including reasoning.

Anthropic:

One note from the Anthropic page worth knowing: Sonnet 5 launched at $2 in and $10 out as introductory pricing that was scheduled to rise to $3 and $15 on September 1, 2026. That increase was cancelled and the launch price is now the standard price.

OpenAI:

Two OpenAI-specific rules change these numbers. Fast mode doubles the price. And on the GPT-6 and GPT-5.6 families, a request whose input exceeds 272,000 tokens is repriced for the whole request, so GPT-6 Astra jumps to $20 in and $75 out past that line. Anthropic's page says the opposite: Claude 4.6 and later include the full one-million-token window at the standard rate, and a 900,000-token request bills at the same per-token price as a 9,000-token one.

For what the new flagship means for store copy specifically, see our GPT-6 Astra for ecommerce content guide. Read across both tables and the market has settled into three tiers. Flagship at $10 in and $50 out, the same at both vendors. Everyday models at $2 to $5 in and $10 to $25 out. Small models at $0.20 to $1 in and $1.20 to $5 out. The gap between the top and bottom tier is roughly fifty times. Most of your savings live in choosing the tier, not in trimming prompts.

Where the money goes in a real request

Numbers on a price sheet are abstract until you see a bill. So on September 4 we ran our own prompts, the ones our app uses to write product descriptions and collection pages, against two OpenAI models with identical inputs and read the usage fields back. Same prompt, same product data, same product image, same day.

Product description for a floral wrap midi dress from a fashion boutique on our platform, with one product image sent at low detail:

Product description for a hand-blown glass vessel from a design store on our platform, same setup:

Collection page for a Casual Dresses collection, text only, eight products in the payload:

Three lessons fall out of these six rows.

The image is the input bill. The text of the product prompt is about 1,300 tokens. GPT-5.4 billed roughly 3,000 more input tokens than GPT-5.6 Sol for the identical request, and the only thing in the request that is not text is the photo. When the payload was text only, both models billed exactly 2,945 input tokens. If you send images, how the model counts them matters more than anything you do to your wording.

Output is where the price per token bites. Even with a cheaper input side, the newer model cost more overall on every run, because it wrote a little more and thought before writing, and thinking is billed as output at the higher rate.

A product description costs three or four cents. That is the honest scale of this. The cost problem in AI content is not the single call. It is the ten thousandth call, the retry nobody noticed, the prompt that quietly grew from 1,300 tokens to 9,000 because someone kept adding instructions.

Nine ways to spend fewer tokens

1. Pick the smallest model that passes your own test

The fifty-times gap between tiers is the biggest lever on the list and the one most teams skip because it requires writing a test. Take twenty real examples of the job, run them on a small model and a flagship, and have a human judge them blind. For classification, extraction, translation, tagging and short structured copy, the small tier usually passes. For long editorial copy where the merchant will read every sentence, it often does not. Route each job to the cheapest model that passed, and keep the test so you can rerun it when a new model ships.

2. Cache the part of the prompt that does not change

Every request you send has a static part (your instructions, your format rules, your examples) and a dynamic part (this product, this customer). Prompt caching lets the vendor store the static prefix and charge you a fraction to reuse it.

On OpenAI it is on by default. Prompts of at least 1,024 visible input tokens on GPT-5.6 and later (2,048 on older models) get cached automatically, cached tokens bill at one tenth of the input rate, and on GPT-5.6 and later a cached prefix stays warm for 30 minutes after its last use. The one rule is structural: put the stable instructions first and the changing content last, because the cache only matches from the start of the prompt.

On Anthropic you opt in with a cache_control field. A five-minute cache write costs 1.25 times the input rate and a one-hour write costs 2 times, and every read costs one tenth, or one fortieth on Fable 5.1. Anthropic's own arithmetic: the five-minute cache pays for itself after one read, the one-hour cache after two.

For a system that writes a hundred product descriptions from the same 1,200-token instruction block, caching turns 120,000 instruction tokens into 12,000. Both vendors say caching stacks with the batch discount.

3. Batch anything nobody is waiting for

Both vendors take 50% off input and output for asynchronous work through their Batch APIs. Bulk product enhancement, nightly SEO audits, tagging a back catalog, regenerating alt text: none of it needs an answer in eight seconds. OpenAI also offers Flex processing at the same half price for lower-priority live requests. The question to ask about every job is whether a human is staring at a spinner. If not, batch it.

4. Cap output, and cap reasoning

Output tokens cost five to six times input tokens on every model in both tables, and reasoning tokens are output. Two settings control this. A maximum output token limit stops a runaway answer. A reasoning effort setting (GPT-6 Astra exposes low, medium, high, xhigh and max) decides how much the model thinks before it writes. A product description does not need maximum effort. A legal summary might. Set effort per job, not per account.

5. Stop sending empty fields

Payloads grow. Someone adds a field, a null rides along forever, and a model reads an empty string as a value to consider. We learned this the hard way and now drop every empty key before a prompt ships. It saves tokens, and it also improves output, because the model no longer wonders what to do with "material: none".

6. Treat images as the expensive input they are

Send one image, not five. Send it at low detail unless the job needs to read fine print. Resize before sending. And if the same image is used across several calls, structure the prompt so it sits in the cached prefix. In our runs above, the image was the single largest line on the input side, and on one model it was three quarters of it.

7. Stay under the long-context cliff

On OpenAI's current models, one token past 272,000 input tokens reprices the entire request at double the input rate and 1.5 times the output rate. If you stuff a whole catalog into a prompt, you can pay twice for every token in it. Chunk the job, or retrieve only the products the task needs, or use a model whose long context is flat priced.

8. Count your tool definitions

Every tool you declare rides in the prompt as tokens on every call. Anthropic publishes the overhead: the tool-use system prompt alone is 286 to 804 tokens depending on model and tool-choice mode, and declaring the computer-use toolset adds about 4,500 input tokens per request. If an agent has forty tools and uses three, you are paying for thirty-seven descriptions on every turn. Give agents the tools the job needs, not the whole box.

9. Put every call in a ledger

Nothing on this list works without measurement. Every call to a model should land in one place with the model, the job that asked for it, the input and output tokens, and the cost at that day's rate. Inside our own platform every LLM, image and video call does exactly that, and the ledger has caught more waste than any prompt review: a retry loop that ran three times, a payload that grew when a field was added, a model that was quietly billing three times the image tokens of its neighbour. Cost you cannot see is cost you cannot cut.

What this means for a Shopify store using AI tools

If you buy AI as a feature inside an app rather than by the token, this guide still matters, because the vendor's token bill sets your price. Ask three questions of any AI tool you pay for. Which model tier does it use for which job, and can it tell you? Does it cache and batch, or is every call a fresh full-price request? And can it show you what one unit of output actually cost to produce? A vendor who can answer those is passing savings through. One who cannot is passing the bill.

Tokens are not going away. The prices in the tables above will be wrong in six months, probably lower for the small tiers and about the same at the top, which is where both vendors have landed on the same number twice in a row. The nine levers do not change. Choose the tier, cache the prefix, batch the background, cap the output, and write everything down.