AI credit pricing has three different exchange rates, and vendors don't publish any of them

AI credit pricing has three different exchange rates

In the past eighteen months, AI credit pricing has gone from a niche, early-startup quirk to a software-industry default. Atlassian, Zendesk, and HubSpot — three vendors that overlap with almost every B2B buyer's stack — have each bolted an AI feature onto an existing subscription and priced it in something called a credit. The word implies a stable, fungible unit. It isn't one. Each company decided, on its own, how many dollars a credit is worth, what action burns one, and what happens when you run out. None of that exchange rate sits on the public pricing page.

That matters beyond vendor-comparison curiosity. Zylo's 2026 SaaS Management Index found that 78% of IT leaders had an unexpected charge tied to a consumption- or AI-based pricing model in the past year, and 61% had to cut a different project to absorb the cost. AI-native app spend at large organisations is up 393% year over year. None of that growth shows up as new tools bought. It shows up as the same tools billing differently than the contract implied.

The fix is duller than the problem deserves: convert every vendor's credit back into a real number before you sign, and check what happens at the edge of your allowance. Below is that conversion, done for three of the credit systems a mid-size B2B company is most likely to actually run into.

Atlassian's Rovo: credits bundled into the seat you already pay for

Atlassian ties Rovo, its AI layer for Jira, Confluence, and Jira Service Management, to a monthly credit allowance that scales with the plan you're already on. According to Atlassian's own usage-limits documentation, a Standard-edition seat gets 25 Rovo credits a month, Premium gets 70, and Enterprise gets 150. A Rovo Chat or Rovo Agent request costs 10 credits; a Deep Research query costs 100. Do the arithmetic and a Standard-tier seat can run roughly two Deep Research queries, or two to three chat requests, before the monthly pool is gone.

The detail that matters most: there is currently no overage billing on core Rovo. Atlassian has committed to 90 days' notice and an explicit opt-in requirement before it starts charging for usage past the included allowance. That's an unusual amount of restraint in this category, and it's worth treating as the bar other vendors in this piece should be measured against, not the default you should assume everyone meets.

Atlassian's separate developer product, Rovo Dev, is priced differently and already has overage billing live: $20 per developer per month including 2,000 credits, with extra usage billed at $0.01 per credit. The gap between free overage today (core Rovo) and billed overage today (Rovo Dev), inside the same company, is itself informative. Overage billing on AI credits isn't a hypothetical feature vendors might add later. It's a switch Atlassian has already built, and already turned on, for at least one of its own product lines.

Zendesk: a credit system with a mandatory toll booth

Zendesk skips the credit abstraction and prices its AI agent by outcome directly: $1.50 per automated resolution at committed volume, or $2.00 pay-as-you-go, on top of a mandatory $50-per-agent-per-month Advanced AI add-on layered over an existing Zendesk plan. Each licensed agent seat carries a small allowance of free automated resolutions, roughly 10 a month, before the per-resolution charge applies.

Run the numbers for a 20-agent support team. The add-on alone costs $1,000 a month before a single ticket gets resolved. The team's 20 seats earn 200 free resolutions a month combined. If the AI resolves 500 tickets, the team pays for 300 extra: $450 at committed pricing, $600 pay-as-you-go, stacked on the $1,000 base. Scale that to a team resolving 3,000 conversations a month and the published math puts the all-in annual cost at roughly $80,000.

This is the cleanest version of the new pricing logic: the vendor is charging for verified value, a resolved ticket, rather than a seat that might use the feature, and that's a defensible idea on its own terms. The catch is that the cost now scales with support volume, which is itself a function of how many customers a company has. A good growth quarter and a bad pricing quarter can look identical on the invoice.

HubSpot: credits that can upgrade your contract without asking

HubSpot prices its Breeze AI features in HubSpot Credits, sold and overage-billed at $0.01 each, or $10 per 1,000. Since 14 April 2026, a resolved Customer Agent conversation consumes 50 credits, which works out to $0.50 a resolution; an unresolved conversation costs nothing. The Prospecting Agent meters differently, at $1.00 per recommended lead. Included allowances scale with plan: Starter gets 500 credits a month, roughly 5 resolved conversations; Professional gets 5,000, roughly 50; Enterprise gets 10,000, roughly 100.

Two structural details push this past simple per-action pricing. Credits reset monthly and unused ones do not roll over, so a quiet month doesn't bank a credit for a busy one. And by default, HubSpot auto-upgrades your account to a higher credit tier the moment you exceed your allowance, locking the account into that higher commitment for the rest of the contract term. You can move back down only at renewal. A single viral marketing push or a bad support week can land an account on a new price tier it didn't choose and can't immediately undo.

Converting credits to dollars is the only way to compare them

Line the three systems up and the comparison only works once everything is translated into the same unit: real dollars per real action.

Three companies, three different decisions about what to meter — seats with headroom, outcomes, conversations — and three different decisions about what happens at the edge of the allowance: nothing yet, a per-unit charge, or an involuntary upgrade. The word credit is doing a lot of work to make these look like the same kind of thing. They aren't.

Why vendors reach for credits instead of a plain per-action price

Credits exist because a plain per-action price has a problem most SaaS companies would rather not advertise: it reveals the vendor's margin. The pricing-strategy blog SoftwarePricing.com lays out the mechanics in its breakdown of what it calls the six fatal flaws of credit-based AI pricing. A credit that converts cleanly to "$1 per 1,000 tokens" invites a customer to look up the foundation model’s published API rate and negotiate the vendor down toward cost. An opaque, multi-layered credit system is much harder to benchmark against anything, which is precisely the point.

The same dynamic explains why credits are usually sold prepaid and drawn down quietly across the contract term rather than billed as they're used. Spending friction that would normally show up every month instead shows up once, at renewal, as a single number nobody budgeted for. SoftwarePricing.com cites a case where exactly this pattern produced a $600,000 surprise invoice large enough to end the customer relationship outright.

None of this makes credits a scam. It makes them a deferred, partially hidden pricing decision, set unilaterally by the party that also controls the exchange rate, and revisited only when that party decides the rate needs to change.

What Cursor's 2025 rollout taught the rest of the market

Atlassian, Zendesk, and HubSpot have all, so far, avoided the failure mode that defines this category: changing the exchange rate after customers have already budgeted around it. Cursor did not.

In June 2025, the AI coding tool replaced its Pro plan's fixed allotment of 500 fast model requests a month with a credit pool worth $20 of usage at the model provider's current API rates. For lighter users, not much changed. For developers running long, agentic coding sessions, the same workflows that used to be flatly included started drawing down a wallet fast. One widely cited account on Hacker News described roughly $350 a week in overage, an effective monthly cost north of $1,400 for work that had cost a flat $20 the month before. Cursor's own account of what went wrong is unusually direct: the company says it communicated the change poorly and underestimated how differently it would land across its user base. It apologised on 4 July 2025 and refunded charges incurred between mid-June and early July.

The market's response wasn't to abandon usage-based pricing. It was to put a floor under it. Kyle Poyar's 2026 State of B2B Monetization survey, run across more than 230 software companies, found that 37% now use a hybrid model — a subscription base plus a usage layer on top — as their primary pricing structure, up from 25% a year earlier. Hybrid pricing keeps the upside of charging more for customers who use more, while keeping the worst-case bill bounded by a number both sides agreed to in advance.

Three questions to ask before you sign, or before you ship your own credit system

If you're buying: ask the vendor, in writing, what one credit costs in dollars, what specific action consumes one, and what happens the moment you exceed your allowance — a pause, a per-unit bill, or an automatic upgrade you can't undo until renewal. Model the worst-case month assuming every seat you're buying is used at capacity, not the usage pattern the sales deck assumes.

If you're building a credit system of your own: publish the exchange rate. Atlassian's restraint on overage billing, and its 90-day-notice commitment before changing that, is the detail every other vendor here should be measured against. A credit that converts cleanly to a dollar figure a customer can check is one that customer will still trust at the next renewal. One that doesn't is a bill shock with a twelve-month fuse.

The vendors that get this right over the next two years probably won't be the ones with the cleverest credit mechanics. They'll be the ones who made the mechanics boring enough that nobody had to write an article translating them back into dollars.