AI pricing explained: subscriptions, credits and token bills

AI tools charge in one of three ways, and the shape matters more than the headline number. A flat subscription is a fixed monthly fee with usage caps that are rarely published precisely. A credit system hands you a pool of units that drains at different rates depending on what you ask for. Per token billing charges for every chunk of text going in and coming out, with no ceiling at all. Most people on a subscription never need to think about tokens. Anyone building on an API thinks about almost nothing else. The arithmetic below shows where a bill actually comes from, and the largest line is usually not the one you expected.

The three ways AI tools charge

Flat subscription. One monthly fee, one model tier, and limits that are the real product. Caps usually run on a rolling window of a few hours rather than a calendar day, with separate allowances for image generation and the slower reasoning modes. You cannot overspend, which is the main attraction. You also cannot carry unused capacity forward, and the exact numbers are often vague until you hit them. Whether that fee earns its place is a separate question, covered in what a paid plan actually buys you.

Credits. A pool of units where each action costs a different amount: a short chat message might be one unit, a long document analysis twenty, an image several. Credits are hard to compare across products by design, since one company's credit is not another's. Three things decide whether a credit plan is good value: the conversion table, whether unused credits expire at the end of the month, and whether heavy actions are priced honestly or used to push you upward.

Per token. The raw model, billed by the chunk of text, quoted per million tokens, with input cheaper than output. No cap, which is both the appeal and the risk. This is what sits underneath most other products, including many of the subscriptions, and it is the only model where your bill scales exactly with what you do.

Plenty of products blend these: a subscription that includes a monthly credit allowance, then charges per token for overage. When that is the structure, find the overage rate before you find the monthly fee.

What a token bill is made of

A token is a chunk of text, usually a common word or part of a longer one. English runs about 1.3 tokens per word, so 1,000 tokens is roughly 750 words. Code, names and most non English text fragment into more tokens for the same meaning. The full mechanics are in what tokens are and how they are counted.

Input and output are priced separately. Output usually costs several times more per token than input, because each output token is generated one at a time and costs real compute. That ratio misleads people, though, because the volumes are lopsided. Document work is overwhelmingly input. A chat that reads a 12 page report and writes four paragraphs sends ten times more tokens than it produces.

The piece nobody expects is the resend. A model has no memory between turns, so the entire conversation so far is sent again as input on every single turn. Ten turns of chat is not ten small charges, it is a charge that grows on each one.

A realistic month, counted out

Take one person doing ordinary knowledge work on per token billing.

Job one: summarizing reports. Twenty reports a month, each about twelve pages, so roughly 6,000 words or 8,000 tokens. Add 300 tokens of instructions each time, and ask for a 500 word summary, about 700 output tokens. That is 20 times 8,300, or 166,000 input tokens, and 14,000 output tokens. Modest.

Job two: the daily working chat. Twenty working days, one conversation each, twelve turns per conversation, with each turn adding around 600 tokens to the transcript once your question and the reply are counted. The naive guess is 7,200 tokens per day. The real figure is 600 multiplied by 1 plus 2 plus 3 and so on up to 12, which is 600 times 78, or 46,800 input tokens per day. Over twenty days that is 936,000 input tokens, plus roughly 72,000 output tokens.

Total for the month: about 1.1 million input tokens and 86,000 output tokens. Note what happened. The document work you thought of as the expensive part is 15 percent of the input. The chat you thought of as free is the rest, and almost all of that is the same text being sent over and over.

Converting to money is one multiplication, but the ratio is what to carry with you. If output costs five times input per token, then 86,000 output tokens cost the same as about 430,000 input tokens, so output is roughly a quarter of this bill despite being a fourteenth of the volume. Long inputs and long outputs need watching for different reasons.

Turning your own work into token estimates

Use this to size a job before you run it. The numbers are English prose unless stated, and they are estimates rather than promises.

Retrieval systems add a line you never see. When a tool searches your documents and attaches the relevant passages to your question, those passages are input tokens too, often several thousand per question. That is one of the quiet cost differences between the approaches in retrieval versus fine tuning.

Why a limit appears in the middle of a task

Three separate mechanisms all get called the limit, and they behave differently, so the fix depends on which one you hit.

• A plan quota. Your allowance of messages or heavy actions, usually on a rolling few hour window. Heavy morning use is exactly what produces an afternoon wall. It resets by waiting.
• A rate limit. Requests or tokens per minute, there to protect shared infrastructure rather than to bill you. It usually clears in seconds and is the one worth retrying automatically.
• The context limit. Not billing at all. This is how much the model can hold in view at once, and hitting it means the tool either refuses or silently trims. That second behavior is why summaries sometimes skip a chapter, as explained in why an AI forgets the start of a long chat.

Long documents, image generation and the slower reasoning modes drain a plan quota far faster than short chat messages, which is why the wall tends to arrive on the day you are doing the work that matters.

The costs that are not on the price page

Retries. Every regeneration re-bills the full input, not just the new output. If one answer in four gets regenerated, add a quarter to your estimate.

Reasoning tokens. Models that think before answering produce internal working that is billed as output even when you never see it. A reply with 300 visible words can carry several thousand tokens of hidden reasoning behind it, which is why the thinking modes come with their own separate caps on subscription plans.

Agent loops. A tool that takes multiple steps on your behalf resends the accumulated context at every step, and a ten step task carries the same triangular growth as a ten turn chat. Costs from an assistant that takes actions are hard to predict for exactly this reason.

Failed formats. When you need strict, machine readable output and it comes back malformed, you pay for the bad attempt and the good one. Validate the output and retry only the failures rather than the whole batch.

Working out what you would really pay

1. Count one typical week honestly: documents processed, chat turns, images generated. Multiply by four, not by five, because nobody works every week.
2. Convert to tokens with the table above, then apply the resend multiplier to anything conversational.
3. Add a quarter for retries and a separate line for reasoning modes if you use them.
4. Compare that with a subscription. If the subscription is cheaper and you never hit the cap, take it. The predictability is worth something on its own.
5. Structure prompts to be cacheable: fixed material such as instructions, examples and the document goes first, the varying question goes last. Providers commonly discount input they have already processed, and that discount only applies when the unchanged part is at the start.
6. Send work in batches when nothing is waiting on the answer. Delayed processing is usually cheaper than immediate.
7. Match the model to the job. Classification, extraction and tidying do not need the strongest model, and routing routine work to a smaller one is the single largest saving available to most people.

If a large share of your volume turns out to be routine, repetitive and not especially hard, there is a fourth pricing model worth knowing about: paying once for hardware and nothing per token afterward. Running a model on your own device covers what that realistically gets you, and where it falls short.