Monthly AI budget, how to estimate it accurately

A monthly AI budget can be estimated in one line. You take the cost of one action, multiply it by the number of actions per month and add a safety margin. All the difficulty lies in the first two terms. It's a matter of method rather than math. You have to define your unit of use, count its tokens and project your volume honestly. The full method follows, with three budgets worked out down to the first cent.

The stakes go beyond bookkeeping curiosity. A budget set in advance can be compared with the plans on the market, defended in front of a client or a partner and monitored without stress. Without one, people overpay out of caution and buy plans sized for whatever might happen. Ten minutes of math against months of spending by guesswork is an easy call. The method wins every time.

01Key takeaways

• The formula is budget = cost per action × monthly actions × 1.3, with the 30% margin covering iterations and surprises.
• Define the unit of use first. It's an article, a conversation or an analysis, never tokens in the abstract.
• Three typical profiles frame the volumes (light, medium and heavy) and you can place yours in five minutes.
• The worked examples give $0.33 a month for a solo writer, a few dollars for a small business assistant and a few dozen for an agency.
• A spending cap set with the provider locks the budget in place, so the estimate becomes a guarantee.

02The budget formula in one line

Monthly budget = (input tokens × input price + output tokens × output price) ÷ 1,000,000. You multiply that by the number of actions in the month, then by 1.3 for the margin. Prices come from the official price lists, tokens are counted with two ratios and actions are projected by profile. That makes three ingredients and three steps, with nothing hidden anywhere.

This budget transparency is one of the most concrete benefits of the AI without subscription framework. When you pay per use, the budget is BUILT from what really happens instead of being imposed by a plan tier. The next sections turn the formula into a road map you can follow step by step, from the unit of use to the spending cap.

03Step 1, define your unit of use

The first decision is also the most neglected, which is choosing what you count. The right unit is the one your work runs on. That's the article for a blogger, the conversation for a customer assistant and the analyzed file for a consultant. The whole budget flows from it. Estimates that miss almost always start by counting abstract tokens instead of business units.

Next, list what one unit really contains. An article includes its brief, its iterations and its metadata. A conversation includes the history sent back each time. An analysis includes the document you send. Being honest about that scope is what makes the final budget accurate, far more than prices recorded to the cent. A complete scope with rounded prices always beats a truncated scope with exact prices.

04Step 2, count the tokens in one unit

Two ratios are enough. English runs at about 0.75 words per token and French at a little over half a word per token. A 1,000-word article in French therefore weighs about 1,700 output tokens, while the same length in English comes closer to 1,330, so the budgets below err on the safe side. A careful brief takes 400 to 800 input tokens. For an exact count on a given text, the official tokenizers do the job. For the full method on a single call, costing each call one by one is the ten-minute groundwork.

Conversations are a special case, because each turn sends the history back as input and the tenth exchange pays for the nine before it. Count the cumulative input of a typical conversation rather than that of a single message. It's the most common gap between an estimated budget and the real statement for assistant use. The cache on repeated input fortunately absorbs a good share of it.

05Step 3, project your monthly volume by profile

Three profiles frame reality. Light is occasional use, with ten to thirty actions a month, like the weekend blogger or a rarely used assistant. Its budget stays under one euro whatever the model. Medium is regular production, with fifty to two hundred actions, like the active creator or the small team. Heavy is daily automation, with a thousand actions or more, like the multi-site publisher or the high-traffic assistant. It's the only profile where the choice of model really weighs on the final bill.

Project your busiest foreseeable month, not the average month you hope for. That's the month the budget has to hold. It's also the one you derive the safety cap from. Quiet months will simply cost less, which is the whole beauty of paying per use rather than paying fixed amounts. Still, note the seasonality of your work, because it sharpens the projection from one year to the next.

06Step 4, apply today's real prices

All that remains is to plug in the price lists. On the official figures of October 16, 2026, output costs $1.20 per million tokens on DeepSeek V4.1-Flash at peak hours, $4.50 on GPT-5.4-mini and $15 on Claude Sonnet 4.6. The full range of twenty-two models sits in the providers' price lists side by side. Pick a model for each task rather than one model on principle.

Think about currency and tax when you top up. Price lists are in dollars, you pay in your own currency at the day's rate and VAT depends on your status. At the scale of the budgets worked out here, the effect is cosmetic. For company accounts, note the rate at each credit top-up. The provider's statement does the rest with an exact history of every top-up.

07Three budgets in figures, worked out end to end

First comes the solo writer, with ten 1,000-word articles a month and one iteration each on GPT-5.4-mini. Twenty generations at $0.0083 make $0.17. Counting metadata and rewording generously at 50% more gives $0.25. With the 30% margin that comes to $0.33 a month, which you can round up to one euro to sleep soundly. That's the AI budget that was causing worry, a few small coins a month with a generous margin, for ten published articles.

Next is the FAQ assistant of a small business, with a thousand conversations a month of six turns each. Each one uses about 4,000 cumulative input tokens and 1,200 output tokens on a budget model such as Haiku. The math gives $10 a month and $13 with the margin. Last comes the agency, with two hundred pieces of content a month for several clients, routed between a floor model and a mid-tier one. That's around $8 of API spend. It comes to about twenty dollars once you include the margin and images billed per piece. Separate keys split it by client, so it can be billed back to the cent from the provider's statements.

08The pocket estimator, ready to copy

1. Line 1, your unit of use and what it really contains (brief, iterations and metadata included).
2. Line 2, the input and output tokens of that unit, from the step 2 ratios or the official tokenizer.
3. Line 3, the input and output prices of the chosen model, taken from today's official price list.
4. Line 4, the unit cost from the formula, multiplied by the monthly volume of your profile.
5. Line 5, times 1.3 for the margin gives the budget. Set the provider cap at that level and date the calculation.

Five lines in a spreadsheet give you a personal version of the online calculators, calibrated on YOUR assumptions rather than default values. Keep the sheet. Every new project can be costed in it in two minutes. The history of your calculations tells the story of how your practice has changed over time.

09The four mistakes that throw budgets off

The first mistake is a vague unit. Budgeting for "AI" without defining the action leads to rough guesses that are always wrong in the painful direction. The second is a truncated scope. Counting the article without its iterations or its brief underestimates by a third, which is the classic gap of the first few months.

The third mistake is forgetting the history in conversational use. Counting the message instead of the cumulative conversation cuts the real budget by three on paper, until the statement corrects it. The fourth is the single model. Budgeting everything at flagship prices inflates the figure for nothing, while budgeting everything at floor prices leaves you exposed. Budget on the routing you really use, task by task.

None of these mistakes is fatal, since the first month's statement reveals all of them. Knowing them in advance still saves you a month of drift and the false conclusion that usage can't be predicted. Usage is predictable. Sloppy assumptions aren't. The statement always condemns them, never the formula.

10The safety margin, 30% and a lock

The 30% margin isn't a quirk. It absorbs unplanned iterations, briefs that grow longer and unusual months, which is exactly the gap that real statements show against honest estimates. If the gap keeps going above 30%, you don't need a bigger margin but a review of your assumptions. Go back to the unit of use and its scope, because the culprit is almost always hiding there.

The lock is what changes the status of the budget. A spending limit set with the provider turns the estimate into a guarantee, so any drift becomes an alert rather than a bill. It takes a minute and lines up naturally with your budget plus margin. Setting a spending cap gives the steps. With the budget calculated and the cap in place, the fear of pay-per-use billing is gone.

11Keeping to the budget over time, without a spreadsheet

Tracking takes a two-minute monthly check. You open the provider's statement, compare it with the budget and note the gap. The dashboards break spending down by day and by model, which is enough to spot a greedy automation or a bloated template right away. Reading your usage as it happens reviews the options provider by provider.

When the gap grows, the levers are ready. The automatic discount on repeated input trims recurring templates, batch processing trims deferred generations and the toolbox of cost levers stacks up the rest. Most of the budgets shown here go down on their own from quarter to quarter, carried by falling prices. Tracking mostly serves to notice it and decide what to do with the difference.

12Solo budget or team budget, what changes

Working solo, the budget described here is enough as it stands, with one calculation, one cap and one statement. A team adds two refinements. One is splitting costs, with one key or one project per person or per client, so the statement can be read without a spreadsheet. The other is governance, which decides the allowed models for each type of task so that routing doesn't depend on everyone's habits.

The cap then applies key by key. Each member or each client carries a limit calibrated on their share of the budget. The limits add up to the total budget. The setup takes an hour for the first person added, then ten minutes for each one after. It prevents unpleasant end-of-month conversations better than any usage charter pinned up in the break room.

13How the budget changes over time

The structural news is good. With the same practice, the budgets shown here go down whenever a provider cuts its price list. Your two-minute monthly check will show it coming. The only decision left will be to pocket the drop or reinvest it in quality by moving up a tier on the content that matters. Both choices make sense. What counts is deciding with the facts in hand.

Volume increases show up in your work before they appear on the statement, with a new site, a new client or a new automation. The healthy habit is to redo the five lines of the estimator whenever the scope changes and to adjust the cap at the same time. The budget then stays a steering tool rather than a relic of the first month that nobody looks at anymore.

14A special case, AI resold inside a service

If your AI powers a service you bill for, such as an app, a service contract or your own subscription, the budget changes nature and becomes a cost of goods. Set a ceiling for the AI cost as a share of the billed price, so the margin survives your other costs. The per-unit calculation described here applies directly to each unit sold.

The extra discipline is to recalculate whenever a price list or a model changes, because the cost of goods is a living number. That's where BYOK shows its structural advantage. The cost is public and traceable to the cent, while a dependence on opaque credits would make the margin impossible to calculate. You can only steer well what you can read.

15Comparing your budget with the market's plans, the final exercise

A budget once set serves one purpose right away, which is a fair comparison. Put your $0.33 a month (or your $13) against the plans competing for the same work, whether tool plans, credit packs or subscriptions. The decision is then made on comparable numbers, with the same volume of work on both sides, rather than on impressions.

The exercise often reveals surprises in both directions. A plan judged expensive turns out to be justified by platform features you really use, while a so-called budget pack costs ten times the calculated budget. Without a budget those truths stay invisible. With one, every line of AI spending has to say exactly what it's being compared against. That's the quiet power of ten minutes of method. Repeat the exercise once a year, since plans move as much as price lists do.

16Recurring budget or one-off project, two separate calculations

The method targets the recurring budget, the one that comes back every month with your work. One-off projects are costed separately. Migrating two hundred old articles or generating a catalog is budgeted as a single envelope, total volume times unit cost plus the same 30% margin. The calculation is identical and only the timing changes. The provider's statement also tells the two apart at a glance if the project carries its own key.

This split avoids the classic false alarm. The project month seems to blow the budget, when it really adds up two planned expenses. One more line in the estimator, dated and named, is enough to keep the story readable. The following month falls back to its normal pace on its own, with no change to the cap needed.

17From a set budget to the full setup

With the budget in hand, three extra steps close the loop. You can place your figures among the orders of magnitude observed on real profiles. You can review the choice between plans if a flat plan is still in the running. Then you can get the API key from the provider you picked, which takes five minutes per path.

OpenTahr fits this method, with a free account, your own key and list prices. Usage stays visible at your provider as it happens and the provider's caps act as a safety net. The monthly AI budget is then what it should be, a number you chose rather than a surprise that lands on you at the end of the month.

18Frequently asked questions

How can you estimate a monthly AI budget quickly?

Use the formula. Take the cost of a typical action, calculated from the official price list of the chosen model, multiply it by the projected monthly volume and add a 30% margin. The full exercise takes ten minutes with the ratios given here. The statement from your first real month then calibrates the ones that follow.

How much should you plan per month for personal use?

For a freelancer producing content, count from a few dozen cents to a few euros depending on the model and the volume, margin included. The worked examples give $0.33 for ten careful articles a month. A first top-up of five dollars usually covers several months of that kind of use.

How do you avoid blowing your AI budget?

Three combined habits do it. Set the budget with the method, lock a spending cap with the provider at the level of the budget plus margin and keep the two-minute monthly check on the statement. With that trio, overspending becomes technically impossible and any gap shows up before it costs anything.

What share of the selling price should AI cost represent?

For a service you bill, set a target share of the selling price for the AI cost, with the rest covering your other costs and the margin. The per-unit calculation described here gives that ratio directly. Watch it whenever a price list or your volume changes.

Should the budget include images and audio?

Yes, on their own price lists. Images are billed per piece, a few cents at most in everyday use. Audio is billed by duration. Add them as extra units of use in the same formula. For a normally illustrated site they add one to two euros a month, margin included.

Disclosure. Prices were recorded on October 16, 2026 from the official price lists cited in the sources. The example budgets are recalculated from those lists with the assumptions stated. You can check them on a calculator. The cost target for billed services comes from published industry practice and varies by trade. None of the providers mentioned pays us.

19Sources

1. OpenAI, official API price list
2. Anthropic, official Claude API price list
3. DeepSeek, official API price list