Tools & Buying
Understanding AI Pricing Models: Seats, Tokens, Credits and Actions
By Jim Vernon, Editor, AI Intelligence International · Published 1 February 2026 · Reviewed against our editorial standards · About the author
Vendors price AI in deliberately incomparable units. Normalising them is the single most useful thing a buyer can do before a negotiation.
There are four common structures, each shifting risk between you and the vendor in a different direction.
Key takeaways
- Per seat: Predictable and easy to budget, but you pay for occasional users at the same rate as heavy ones, and cost scales with headcount rather than value.
- Per token or per unit of text: Closely tracks actual consumption and rewards efficient design, but is hard to forecast and can surprise you when usage grows.
- Credits: Credits obscure the underlying economics, often expire, and vary in what they buy between features.
- Per action or outcome: Charging per resolved ticket, per document processed or per successful outcome aligns incentives well and is the easiest to justify internally.
Per seat
Predictable and easy to budget, but you pay for occasional users at the same rate as heavy ones, and cost scales with headcount rather than value.
Best for tools used daily by a defined group. Worst for tools with a long tail of light users.
Negotiate for tiered or floating seats where usage is uneven.
Per token or per unit of text
Closely tracks actual consumption and rewards efficient design, but is hard to forecast and can surprise you when usage grows.
Understand that both input and output count, and that retrieved context and long system prompts inflate input substantially.
Convert to a cost per completed task before comparing anything. Token prices alone are meaningless across different workflows.
Credits
Credits obscure the underlying economics, often expire, and vary in what they buy between features. This is a pricing structure designed to be difficult to compare.
Ask for the credit cost of your three most common actions and compute an effective price. Then ask what happens when credits run out mid-cycle.
Avoid large prepaid credit commitments in year one.
Per action or outcome
Charging per resolved ticket, per document processed or per successful outcome aligns incentives well and is the easiest to justify internally.
The risk is definitional: what counts as resolved, and who decides. Get the definition in the contract with examples.
Where available and honestly defined, this is usually the buyer-friendly option.
Normalising for comparison
Build one table: your monthly volume of each task, the cost per task under each vendor's structure, and the cost at double volume.
The doubling column is what reveals the risk. Vendors that look cheapest at current volume are frequently the most expensive at success, and that is precisely the scenario you are buying for.
Worked comparison across three pricing shapes
Take one workflow — summarising two hundred support tickets a month — and price it three ways. Per seat: two seats at thirty each is sixty, fixed, regardless of whether you process two hundred or two thousand.
Per token: two hundred tickets averaging a thousand words in and three hundred out lands in single-digit currency units on a mid-tier model, but scales linearly and doubles if volume doubles.
Per credit: the same work might be quoted at four hundred credits, and the honest answer is that you cannot compare it until you convert credits to work units. Ask the vendor how many credits one representative task consumes, and get the answer in writing.
The clauses that change the real price
Rollover and expiry decide whether unused capacity is value or waste. Monthly expiry on a seasonal business is a quiet surcharge on your quiet months.
Overage behaviour matters more than headline rate: some vendors throttle, some bill automatically, some block. Automatic overage billing without a cap is the single most common source of surprise invoices in AI tooling.
Finally, check what a 'unit' includes. Retries, failed generations, system prompt tokens and retrieved context often count, and long retrieval contexts can double an estimate built from user input alone.
Normalise everything to cost per completed task
Convert every quote into the same denominator: what does one finished, acceptable output cost, including retries and the human review time it needs? That number is comparable across pricing shapes; monthly totals are not.
Model three volumes — expected, double, and half — before signing. Pricing that is excellent at your current volume and punitive at double is a growth tax you will notice at exactly the wrong moment.
Questions to send before you sign
Ask five things in writing: what counts as a billable unit, whether retries and system context are billed, what happens at the limit, whether unused capacity rolls over, and what notice period applies to price changes.
Written answers are the point. Sales calls produce reassurance; the contract produces invoices, and only one of those is enforceable when the bill arrives.
If a vendor cannot state how many credits a representative task consumes, treat the pricing as unquantified and price the risk accordingly — usually by starting monthly with a hard spending cap.
Frequently asked questions
Which model is safest for a first deployment?
Per seat or per action with a monthly term. Both cap the downside while you learn your real usage.
How do I estimate token usage?
Measure a representative task, including system prompt and retrieved context, then multiply by realistic volume and add a margin for retries.
Are annual discounts worth it?
Only where usage is proven. In year one, flexibility is generally worth more than the discount percentage.
Can I negotiate?
Usually yes on term length, ramp periods and overage terms, more easily than on headline price.
Which pricing model is safest?
Per seat for predictable budgeting, usage-based for genuinely variable workloads. Credits are the least transparent and need a stated conversion to real tasks.
Should I take the annual discount?
Only after a full billing cycle on monthly. Annual commitments on a tool you have used for three weeks are how audits find dead subscriptions.