Credits or Seats: How AI Dev Tools Actually Charge
Seat pricing broke when every request started costing the vendor money. A survey of what the market charges for now — agent compute, credits, per-action — and how to read a bill before you sign it.
Software pricing used to be simple because software costs were simple: build it once, serve it a million times, charge per seat. AI broke that. Every request costs the vendor real money, so the seat model — which promises unlimited use for a fixed fee — becomes a bet the vendor loses whenever a customer is enthusiastic.
The market has spent two years working out what to do instead. Here is where it has landed, and how to read a bill you are about to sign.
What the market actually charges for
Four shapes dominate, and they meter different things.
Agent compute. Devin prices in ACUs — Agent Compute Units — a normalised measure bundling VM time, model inference and bandwidth, with one ACU representing roughly fifteen minutes of autonomous work. The published rate is $2.25 per ACU on the entry plan, with no free tier. This is the most honest shape when the vendor's own agents are doing the expensive part: you are billed for their compute, plainly.
Credits as a wallet. Replit's entry plan is $20 a month and includes $20 of usage credits, moving to pay-as-you-go beyond it. Read that carefully: the subscription is essentially a deposit plus platform access, not a markup on tokens. Margin arrives later, with volume.
Credits as an allowance. Lovable gives five credits a day free, capped at thirty a month, and $25 a month buys a hundred, with unused plan credits rolling over for a limited window. Here credits are a unit of product action rather than a currency.
Per-unit-of-work. Greptile charges $30 per seat including fifty code reviews, then $1 per additional review, with a free tier of fifty reviews a month. A hybrid: a seat gets you access, the meter runs on the thing that costs them money.
The pattern underneath all four: vendors meter whatever they themselves pay for. Where that is agent compute, they sell compute. Where it is a discrete action, they sell actions.
The free tier tells you the cost structure
Look at what a vendor gives away and you can infer what their marginal cost is.
Devin gives away nothing, because fifteen minutes of autonomous agent work is genuinely expensive. Lovable and Replit drip small daily allowances, which is affordable when each action is cheap and valuable when it builds a habit. Greptile hands out fifty reviews a month, which tells you a review costs them well under a dollar.
Notice what nobody does: a large one-off grant with nothing following it. Every player in this market drips on a recurring schedule. The reason is behavioural rather than financial — a single grant gets burned in one evening and the person never comes back, while a small recurring allowance brings them back on a cadence.
If you see a big one-time bonus and no recurrence, you are looking at a company optimising for signup numbers rather than for retention.
Reading a bill before you sign it
Four questions cut through most of the marketing.
What exactly consumes a unit? Get this in writing, at the level of "connecting a repository consumes X, asking a question consumes Y". Vague answers here become surprise invoices, and a surprise invoice is the most common reason teams abandon a tool they otherwise liked.
Can it overspend? Ask whether units are drawn down as the work happens or estimated up front. Metered-as-spent means the work simply stops when the balance does — no overrun is possible. Estimated up front means somebody eventually gets a bill they did not authorise.
What happens when the balance runs out mid-task? The good answer is that the work pauses where it is and resumes on top-up. The bad answer is that it rolls silently into overage.
Do units expire? Plan allowances usually reset monthly; purchased credits usually last longer. Both are defensible. What matters is that it is stated, because unstated expiry is where the disputes are.
The question worth asking about metering
Here is the test that reveals the most about a vendor's economics: what do they refuse to meter?
If a tool meters every question you ask it, that tells you a question is expensive for them to serve — and it also means the tool is charging for the behaviour that creates the value. Teams that ration questions ask fewer of them, learn less, and quietly stop using the thing.
If a tool leaves questions unmetered, one of two things is true. Either they have not done the arithmetic, or answering genuinely costs them little. The way to tell them apart is to ask how much context an answer consumes. In a controlled test, the range across tools was 8.4K tokens per answer at one end and 42.7K at the other — a fivefold spread. At the low end, unmetered questions are an affordable commitment. At the high end, they are a promise that has to be walked back.
That is why the cheapest-per-answer product can make an offer its competitors cannot match, and why "unlimited" in this market should be read as a claim about the vendor's cost structure rather than about their generosity.
Choosing for your own team
Match the model to your usage shape, not to the headline number.
- Bursty, one-off work — indexing a repository, a migration, a one-time analysis. Prefer credits you can buy as needed. A subscription sized for the burst is idle the rest of the month.
- Steady daily use — an assistant your team leans on constantly. A subscription with a generous included allowance beats pay-as-you-go, because per-request billing makes people hesitate and hesitation defeats the tool.
- Agents in production — the only number that matters is cost per completed customer outcome, not cost per call. Model it before signing; per-call pricing that looks cheap can be expensive per outcome if the agent needs many calls to finish one thing.
- Regulated or large estates — take the contracted plan. Predictability is worth more than the discount you would win by metering.
And whichever you pick, instrument it yourself from day one. Track spend against work completed, not against calls made. That ratio is the only one that tells you whether the tool is earning its place, and it is the number every pricing page is designed not to give you.
Put a number on your own workflow
Bring us the recurring workflow where AI still needs expensive people to supervise, review and correct. We baseline what it costs and put the target in writing before we build. Or connect a repository and see it on your own code first — 1,000 credits free, no card.
Build My Business Case → Start free →Frequently asked questions
Why are AI development tools moving away from per-seat pricing?
Because every request costs the vendor real money. A seat price promises unlimited use for a fixed fee, which the vendor loses on precisely when a customer is enthusiastic. Metering ties revenue to the vendor's own marginal cost, which is why credits, agent-compute units and per-action pricing have replaced flat seats for anything agentic.
What is an ACU in Devin's pricing?
An Agent Compute Unit — a normalised measure of the resources Devin consumes while working, covering VM time, model inference and bandwidth. One ACU is roughly fifteen minutes of autonomous work, priced at $2.25 on the entry plan. It is a clear example of a vendor metering the thing that actually costs them money.
What should I ask a vendor before signing a usage-based contract?
Four things, in writing: exactly what consumes a unit; whether units are drawn down as work happens or estimated in advance; what happens if the balance runs out mid-task; and whether units expire. Vague answers on the first question are the usual source of surprise invoices.
Is "unlimited" ever real in AI tool pricing?
It can be, but read it as a statement about the vendor's cost structure rather than their generosity. If serving an answer costs them little — and the spread across tools is roughly fivefold, from 8.4K tokens per answer to 42.7K — then leaving that action unmetered is affordable arithmetic. If it costs a lot, the promise gets walked back later.
Related articles
The AI Tooling Market in 2026: An Honest Map
Adoption is over at around 90% and proved little. The four positions in the market, what pricing reveals about cost structure, and buying advice for this quarter.
How to Benchmark AI Coding Tools on Your Own Repository
A practical protocol for testing AI code tools yourself: designing the question set, holding the model constant, blind scoring, and the four things worth measuring.
Context Is What You Pay For
Cost per token is a supplier's metric. The number that discriminates between AI tools is how much correct reasoning you get per unit of context — and the spread across tools is elevenfold.