The real cost of AI when you pay per use

The real cost of AI billed per use fits in a sentence. A freelancer who produces content with their own API key usually pays a few cents a month, rarely more than a few euros. Heavy team use stays under a few dozen euros. These orders of magnitude surprise people, because the prices shown elsewhere (plans or credits) almost never relate to the providers' own price lists.

All the figures that follow were recorded on October 16, 2026 from the official pages of OpenAI, Anthropic, Google and Mistral. They are then turned into worked scenarios, from the solo blogger to the agency, so you can place your own usage before you reach for your credit card. Each scenario states its volume assumptions, which means you can redo the math with your own numbers.

01Key takeaways

• AI APIs bill per million tokens. In October 2026, light models cost less than $5 per million output tokens, while high-end models cost between $15 and $50.
• A full blog article generated through the API costs about one cent on a light model and less than five cents on a high-end one.
• Twenty articles a month cost less than one euro. Three hundred articles come to around ten euros on a quality model, with the detailed figures in the scenarios further down.
• The cost to watch is how your tool counts, not the token price. A platform that resells each action as credits multiplies the cost by ten to a hundred.
• With a spending cap set on day one, the bill can't run away from you, because the limit cuts off requests.

02The short answer, then the method

If you want a single benchmark, count on a few cents to a few euros a month for individual content production. A team that automates seriously will spend a few euros to a few dozen euros. The detail depends on three variables, which are the volume processed, the model you pick and the number of round trips per task. None of them is a black box, since you can read all three day by day on the usage page of your console.

One point before the numbers. These amounts cover the AI alone. If you use a tool on top, its price is added, as a flat plan or as credits depending on how it sells. That second layer is exactly what separates the offers on the market. We come back to it further down.

The range assumes one thing, which is that you pay for AI at the provider's price with your own key. That's the BYOK principle, explained in the guide on how AI without a subscription works. The tool supplies the interface, your key pays the official rates and nobody takes a margin in between. Every calculation that follows is built on that basis.

Monthly ranges found in the scenarios (price lists of October 16, 2026)

Profile	API spend per month
Freelancer, 20 articles a month	under $2
Small team, 1,000 support replies	$1 to $3
Agency, 300 articles a month	$4 to $13
Daily news monitoring automation	$1.50 to $3

The amounts are in dollars, like the providers' price lists. Convert them at the day's rate if you like, since the order of magnitude won't change. Each line is recalculated step by step in the scenarios further down, with every assumption spelled out so you can check it.

03How the API bills you and why it matters

Providers count in tokens, the units the text is cut into. A token is roughly three quarters of an English word. In French it's a little less. A thousand words of generated text therefore come to about 1,700 tokens. If the idea is still fuzzy, the guide on what a token represents explains it with two telling examples.

Two meters instead of one

Every request runs two meters. Input is your instruction plus the context you supply. Output is the generated text. Output is billed much higher, often five to six times the input price, because generating takes more computing than reading. The reasons and practical effects of this gap are covered in the guide on why output costs more.

This structure has a virtue that plans lack. Each task has an exact price known in advance. Take an instruction of 1,000 tokens, an expected output of 2,500 tokens and a public price list. You can do the multiplication in your head. To go from a rate to the cost of one specific task, the step-by-step method sits in the guide to calculate the cost of an API call. Here we apply it straight to the current price lists.

04Official prices in October 2026, provider by provider

The table shows the price lists of the four companies that cover most uses, recorded on October 16, 2026 from their respective pricing pages. Prices are per million tokens, before discounts. A price list always reads in two columns, with what you pay to send text on one side and what you pay to receive it on the other.

API rates per million tokens, official price lists recorded on October 16, 2026

Model	Input	Output
GPT-5.5 (OpenAI)	$5	$30
GPT-5.4 (OpenAI)	$2.50	$15
GPT-5.4-mini (OpenAI)	$0.75	$4.50
GPT-5.4-nano (OpenAI)	$0.20	$1.25
Claude Opus 4.8 (Anthropic)	$5	$25
Claude Sonnet 4.6 (Anthropic)	$3	$15
Claude Haiku 4.5 (Anthropic)	$1	$5
Gemini 3.6 Flash (Google), rate until December 31, 2026	$0.75	$3.75
Mistral Medium 3.5	$1.50	$7.50
Mistral Large 3	$0.50	$1.50
Mistral Small 4	$0.15	$0.60

Two discounts apply almost everywhere. Batch processing halves prices for jobs that can wait a few hours. The cache serves input it has already seen for a tenth of its price. The guide on prompt caching explains how it works. One example shows its weight. A 2,000-token working instruction repeated over fifty articles adds up to 100,000 input tokens, but once it's served from the cache after the first send, it costs the equivalent of only 12,000. Repetitive work is the norm in content production, so the discount is far from trivial. For a finer side-by-side of providers, model by model, see the API price comparison, since OpenAI stays the main thread here.

05The real cost per task, before talking budget

The clearest view comes from going down to the single task. The next table applies the price lists to typical web jobs, with the volume assumptions in parentheses. All amounts are rounded up to the next cent to stay on the safe side.

Cost of one task billed per use, calculated on the price lists of October 16, 2026

Task (volume assumption)	Light model	High-end model
1,500-word blog article (2,500 output tokens)	~$0.02	~$0.05
300-word product page (500 output tokens)	~$0.01	~$0.01
Rewrite of an existing page (2,000 in / 1,500 out)	~$0.01	~$0.03
Summary of a long document (10,000 in / 800 out)	~$0.02	~$0.05
Assistant reply (600 in / 400 out)	~$0.01	~$0.01

The lesson of the table is short. Per unit, almost everything costs one to five cents. The gap between a light model and a high-end model shows in the quality of the first draft, never in a budget difference that would give you pause. Model choices only start to count once you multiply by the monthly volume.

The same table teaches a second thing. The most expensive task isn't always the longest one to generate. Summarizing a long document costs as much as a whole article, because the 10,000 input tokens weigh more than the 800 output tokens. When you estimate a new use case, always look at both columns of the price list before drawing a conclusion.

06Four worked scenarios on these price lists

The assumptions are kept simple on purpose and stated each time. A 1,500-word article represents about 2,500 output tokens and 1,000 input tokens. A short assistant reply comes to about 400 output tokens. You can redo every calculation with your own volumes and see how far your case moves from these.

The freelancer who produces content

Take twenty articles a month on GPT-5.4-mini. That's 50,000 output tokens at $4.50 per million, so 23 cents, plus 2 cents of input. The same volume on Claude Sonnet 4.6, a high-end model, comes to 75 cents of output and 6 cents of input. Either way, the AI line of a serious blog stays under one euro a month. Add a few rewrites and proofreading passes along the way and you still stay under two euros.

Now set that figure against what the same output costs elsewhere. Twenty articles billed in credits, at the common rate of €1.50 per article, come to €30 a month. The work is identical and the model running behind the scenes is sometimes the same. Only the billing circuit changes. Over a year, the gap more than pays for the time spent creating your keys.

The small team that answers its customers

Picture a thousand support replies a month, each with 400 output tokens and a generous 600-token input context. On GPT-5.4-mini that gives $1.80 of output and $0.45 of input, so about $2.30 a month. On GPT-5.4-nano the same volume drops below one dollar. It's the kind of figure that looks wrong until you work it out yourself.

The trap for this profile is the history, not the unit price. If the assistant resends the whole conversation at each turn, the input swells as the exchange goes on. A ten-turn exchange can use five times more input tokens than ten separate questions. Well-built tools trim or summarize the history, so check this point before you plug in a large volume.

The agency in continuous production

Three hundred articles a month on Claude Sonnet 4.6 means 750,000 output tokens at $15 per million, so $11.25, plus about $1 of input. The same volume drops below $4 on GPT-5.4-mini if the task allows it. At this scale, picking the cheapest model for each type of task becomes the first way to save, far ahead of negotiating anything.

Background automation

The last profile is often forgotten. It covers the tasks that run on their own. A news monitor that summarizes thirty articles a day, for example, uses about 300,000 input tokens and 25,000 output tokens daily. On GPT-5.4-nano the full month comes to about $2.80. With batch processing it's half that. This kind of automation would be priced out of reach in credits, yet it turns into budget background noise once you pay the list price.

07Why the figures you read elsewhere differ so much

We cross-checked the guides that rank on the subject. Many of them show prices that bear no relation to the current price list. Three causes keep coming back. Knowing them lets you read any content about the real cost of AI with the right critical eye.

Dates are the first culprit. Price lists move fast, usually downward from one model generation to the next and sometimes sharply. A guide that quotes prices without the date they were recorded or a link to the official list tells you about the past. Mixed-up units come next. Some sites show prices per thousand tokens and others per million. That factor of a thousand wrecks quick comparisons.

A final cause is the most interesting. Many published "AI prices" are really tool prices, in the form of plans or credit packs where the AI share is a minority. Comparing a 15-cent credit with a token at a few millionths of a dollar only makes sense once you bring everything back to the task. The worked scenarios earlier on do exactly that.

Hence the reading rule we apply to ourselves. An AI cost figure is only worth something when it's dated, sourced to the official price list and paired with its volume assumptions. Everything else is a sales argument in one direction or the other. It deserves to be treated as such.

08API versus plans, a question of volume

A chat plan like ChatGPT Plus costs $20 a month. Convert that amount into tokens to get a sense of scale. $20 buys about 4.4 million output tokens on GPT-5.4-mini, the equivalent of more than 1,700 articles of 1,500 words. Nobody produces that in a month. For tool-based content production, the API therefore costs ten to fifty times less than the plan. The plan still makes sense for free-form, heavy conversation, which we detail in the API versus ChatGPT Plus.

The underlying logic is laid out in our analysis of paying for AI per use. As long as your monthly volume stays under several million tokens, per-use billing wins. Beyond that, you compare line by line with a calculator in hand and no dogma. The break-even point takes five minutes to work out with your real volumes. It's the only calculation worth redoing each time a price list changes.

09Against credits, where the gap really widens

The chat plan still isn't the most telling comparison. Many content tools sell credits. Writing an article might cost 10 credits there. That's €1.50 based on a common pack. That same article, as we just calculated, costs one to five cents of API. The gap is far bigger than a few percent, since it's a factor of thirty to a hundred.

There's nothing dishonest in that, because the credit pays for the tool and its margin. The effect is mechanical all the same. At volume you mostly pay the margin, since the underlying AI cost stays in cents no matter what. The setup that reverses this splits the two building blocks, with a capped flat-rate tool that accepts your key on one side and AI at cost price on the other. That's the model we chose for OpenTahr. The subscription covers the workshop and never your AI usage, which stays between you and your provider.

10The side costs to know before you sign up

Per-use billing is easy to follow once you know four details that surprise newcomers. None of them changes the orders of magnitude seen earlier. Each one does explain a small gap between a back-of-the-envelope estimate and the bill shown in the console.

• The minimum prepayment. Most consoles ask you to load a few dollars at the start. Others bill in arrears above a threshold. Nothing blocks you, but the first charge isn't zero.
• Currency exchange. Price lists are in dollars. Depending on your bank, the conversion adds a few percent. Some providers also bill in euros, so look at your console.
• Reasoning models. They use internal thinking tokens, billed as output, that never appear in the final text. For everyday writing, classic models are enough and cost less.
• Conversation history. In assistant-style use, the whole previous exchange goes back in as input at every turn. That's what slowly inflates bills, far more than the last question.

None of these items changes the order of magnitude. The console remains the final judge. Get into the habit to track your API usage every week during the first month. You quickly learn what your normal spend looks like. After that a monthly glance is enough.

11Three mistakes that inflate a bill, all avoidable

When an API bill comes as a surprise, the cause almost always falls into one of three families. Knowing them in advance is worth more than any calculator. The first is using the high-end model by default. Putting the most powerful model on every task multiplies the bill by three to ten with no visible benefit on routine work. The reverse split, light by default and high-end only when justified, gives the same perceived quality for a fraction of the price.

Another mistake is using reasoning models as writing models. Their thinking tokens, billed as output, can exceed the length of the final text. They're valuable on complex problems and out of place for writing a product page. A routine writing job gains nothing from them except a higher line on the invoice.

Automated loops without a cap complete the list. A script that retries endlessly on an error or an agent that keeps feeding itself sends usage up in silence. That's precisely the case a spending limit stops dead. The loop hits the cap, you get the alert and nothing else happens.

12Keeping the bill under control for the long run

Once you know the traps, control takes very little, as long as you set it up on day one instead of after a surprise. Three habits make per-use billing more predictable than a plan. Start with the cap, since every console lets you set a spending limit that's firm and monthly, to configure before the first request. Then comes splitting tasks, with routine work on a light model and the high-end model kept for what justifies it. Last comes the monthly review, five minutes on the usage page to check that spending tracks the work produced.

When volume grows, dedicated techniques take over, such as caching, batches and shorter instructions. They're gathered in the levers that bring the bill down. To turn these orders of magnitude into a clean forecast, the guide on how to build a monthly AI budget helps you avoid being off by half.

13From dollars to euros, reading your bill without surprises

Provider price lists are shown in dollars and before tax. That has two practical effects for a reader in France. One is conversion. Your bank applies its own rate and, depending on the card, a currency fee of a few percent. Some providers also offer billing directly in euros, so check the options in your console. Then there's VAT. Depending on your status (VAT-registered or not) and on the provider's billing circuit, the tax is either added or reverse-charged. The invoice issued by the console is the one that counts. It's also the one your accountant expects.

The right reflex is to think in orders of magnitude rather than to the cent. A scenario worked out at $2 is still a scenario at about €2 once everything is counted. No conversion will turn cents into tens of euros. That's all that matters for the decision.

14Frequently asked questions

How much does the OpenAI API cost per month?

It all depends on volume. For the models in the table, the October 2026 price list runs from $0.20 per million input tokens on GPT-5.4-nano to $30 per million output tokens on GPT-5.5. The newer GPT-6 models extend it from $0.10 (Luna) to $50 (Astra). In practice, an individual content producer sees monthly bills from a few dozen cents to a few euros. The console shows the meter continuously, so you can check your spending pace without waiting for the end of the month. The first step is still to create an OpenAI API key with a low cap and measure one real month.

Is the API cheaper than ChatGPT Plus?

For producing content, almost always. Our scenarios stay in cents where the plan takes $20 a month. For hours of daily conversation, though, the plan caps a usage that would eventually show up token by token. The two products answer different needs, so the right choice depends on how you actually work.

What do 1,000 requests cost?

With short 400-token replies on GPT-5.4-mini, about $2 all in, input context included. On a high-end model like Claude Sonnet 4.6, count around $7. The same request billed in credits on a content platform would run into hundreds of euros, which is the gap to keep in mind.

Can API prices go up?

They can, since nothing is contractual on a standard account. Recent history points the other way, because price lists have dropped at equal capability from one model generation to the next. BYOK protects you by design. Your key isn't a commitment, so you can change model or provider the day a price list no longer suits you.

Are there fixed fees or a hidden subscription?

There aren't. None of the four providers named here charges a monthly fee. You pay for the tokens you use. The only indirect costs are the initial prepayment of a few dollars and any currency conversion. A month without usage is a month at zero, something no plan can offer by design.

Which provider should I pick to pay the least?

On raw price, Mistral Small 4 and GPT-5.4-nano sit at the bottom of the table, while the newer Claude Haiku 5.5 and GPT-6 Luna go lower still at $0.10 and $0.50. On writing quality, the mid-range models compete for the best value AI model. The right move is to test two models on your real tasks. The per-provider key guides get you keys in hand within ten minutes.

Transparency note. All prices quoted here were recorded on October 16, 2026 from the official price lists linked below. The scenarios are arithmetic calculations on those lists, with the volume assumptions stated in the text, so redo them with your own figures. Price lists change regularly.

15Sources

1. OpenAI, API pricing
2. Anthropic, Claude model pricing, caching and batch processing
3. Google, Gemini API pricing
4. Mistral, API pricing
5. OpenAI, ChatGPT plans
6. Mistral, API price list per model (Small 4 at $0.15 and $0.60)