Open your inbox and count. Intercom sells "resolutions." Notion sells "AI blocks." Canva, HubSpot, Salesforce — credits, all of them, proliferating like the loyalty points of a hundred bankrupt airlines. The token, that elegant atomic unit of inference, has been laundered through the marketing department and re-emerged as a "credit." And just like crypto tokens circa 2018, every issuer insists their coin is different, their coin has utility, their coin will hold value. Forgive me if I've seen this film.
How we got here: the SaaS pricing identity crisis
Seat-based pricing was the bedrock of SaaS for two decades, and AI broke it in a single product cycle. When one user with a good agent does the work of ten, charging per seat is charging for a metric that no longer correlates with anything. Vendors panicked, understandably. Their costs, however, are denominated in tokens — a real, metered, thermodynamic quantity that drops 5–10× a year in price. Selling tokens directly would expose that deflation to customers. Can't have that.
Enter the credit. The credit is the SaaS vendor's answer to the same problem crypto faced: how do you price something whose underlying cost basis is volatile, collapsing, and opaque? Crypto's answer was to float a token and let the market discover price. SaaS's answer is to fix an exchange rate, bury it in a pricing page, and quietly devalue it next quarter. One of these is at least honest about its chaos.
Every credit is a different currency, and that's the point
Here's the thing vendors won't put in the deck: a credit at vendor A and a credit at vendor B are not comparable units, by design. One credit might be a frontier reasoning call at $15 per million output tokens; another might be a distilled model at $0.03. A 500× spread hiding behind the same word. When I was evaluating AI add-ons for a client's PLM stack last year, I built a small spreadsheet converting each vendor's credit back into equivalent frontier tokens. Two of the four sales teams could not answer the question "what is a credit, in tokens?" Their own pricing teams hadn't told them. That's not a pricing model. That's a casino chip.
And the devaluation risk is real. Crypto taught us that when issuers control supply, they print. A vendor whose inference costs halve every two months faces a choice: pass the savings through, or pocket them. The credit mechanism exists precisely so you never find out which one they chose.
So which pricing model wins?
Let me take a position, because the hedged answer helps no one. Pure per-token pass-through loses. It's honest, but it outsources cost volatility to the customer, and procurement departments hate volatility more than they hate overpaying. I learned this negotiating enterprise agreements: a CFO will accept a 20% premium for a number she can put in a budget and forget.
Per-seat dies for AI-heavy products, slowly and expensively.
What survives, I'd argue, is outcome-based pricing with a token-metered floor — Intercom's per-resolution model is the early, clumsy version of this. The customer pays for a resolved ticket, a qualified lead, a reviewed contract. The vendor bears the token risk and earns a spread for managing it, like any good intermediary. This aligns incentives beautifully and terrifies vendors whose AI features don't actually work. Which is rather the point.
The alternatives worth watching
- Outcome pricing — pay per resolved case, per generated design, per accepted suggestion. Clean for customers, brutal for vendors with weak models.
- Capacity pricing — rent dedicated inference (a GPU-hour, a reserved throughput tier) and bring your own models. The HPC model, reborn; attractive to any enterprise with volume and a competent platform team.
- Bring-your-own-key — the SaaS layer charges for workflow, you supply the inference from your own Ollama cluster or negotiated API contract. Unbundling, in other words. Vendors hate it. Customers with scale love it.
- Flat-rate with fair-use caps — the consumer answer. Works until power users arrive, then quietly becomes capacity pricing.
The forward look
The tell to watch in 2027 is whether any major vendor publishes a credit-to-token exchange rate and commits to holding it. The first one that does — and survives the next 10× drop in inference costs without devaluing — will have built something crypto never managed: a stablecoin backed by actual productive capacity. Until then, treat every credit balance on your books the way a treasurer treats airline miles. Spend it fast. It was never going to appreciate.
The deeper question is whether pricing follows cost at all, or follows value — and those two curves, cost collapsing while value compounds, are diverging faster than any pricing page can track. Someone will build the business that arbitrages that gap. Will it be your vendor, or your competitor?