The Contract Decays Faster Than You Can Sign It

Software pricing has always assumed a certain stillness. You work out what the thing costs to run, add a margin, sign a three-year agreement, and revisit at renewal. The underlying economics move slowly enough that the arithmetic holds.

That assumption has stopped holding, and the number explaining why is unusually blunt.

Per a16z, cited in Chargebee's playbook on AI-agent pricing, the cost of an LLM at equivalent performance falls roughly 10x per year.

Ten times. Annually. Which means a price set today against today's inference cost is not merely conservative twelve months from now. It is off by an order of magnitude, in a direction that eventually becomes visible to the customer.

Both parties are exposed, in opposite directions

The vendor's problem is the obvious one and it goes the wrong way for the buyer's sympathy. Price against current costs, and margin expands dramatically as inference gets cheaper. That is pleasant right up until a competitor prices against next year's costs and the whole category resets underneath you.

The buyer's problem is more interesting. A multi-year agreement signed at today's rates locks in an input cost that the market will have repriced twice before the term is up. The buyer is not being cheated. They simply agreed to a fixed price for something with a rapidly falling floor, which is a bad trade regardless of anyone's intent.

The naive response is to price below cost and win share. Chargebee flags Cursor's forced move off unlimited plans onto usage caps as the cautionary case. Pricing beneath true infrastructure cost works precisely until usage arrives, at which point the correction is public, abrupt, and lands on customers who planned around the old terms.

Why seats made this worse

Seat-based pricing was a reasonable proxy for value when software mostly sat there waiting to be used. A seat approximated a person, a person approximated usage, and usage approximated value.

Agents break every link in that chain. Chargebee describes it as a three-body problem: the product changes fast, per-user consumption varies enormously, and infrastructure cost is itself volatile. An agent decomposes a goal into many steps, so two users on identical seats can differ by orders of magnitude in what they actually consume. The seat no longer tracks value or cost, and it is not close.

What replaces it, per the same playbook, splits three ways. Outcome-based pricing charges per result, the way Intercom's Fin bills per resolved issue. Action or workflow usage abstracts heterogeneous costs into intuitive units, the way N8N charges per workflow batch or Clay uses a credit burn table. Hybrid pairs a fixed base with variable tiers, bundling included credits into a recurring plan.

Each has a failure mode. Outcome pricing forces a hard definition of "resolution" and under-monetizes failed attempts. Usage pricing invites sticker shock on a spike. Hybrid causes frustration when the packaging is wrong in either direction.

The decision that actually matters

Choosing among those three gets most of the attention. It is the second-order question.

The first-order question is cadence. If the input cost falls by an order of magnitude annually, then any pricing model, chosen well, is still wrong within a year unless somebody is obligated to revisit it.

Chargebee's operational answer is a standing cross-functional pricing committee and continuous recalibration, alongside modeling three cost scenarios before launch: baseline, spike, and supplier price increase. That reads like process overhead until you notice it is the only mechanism that makes a decaying input cost survivable.

For buyers, the equivalent move is contractual rather than organizational. If inference cost is falling 10x a year, a multi-year agreement should say something about that. A repricing trigger, a benchmark clause, a shorter term than instinct suggests. Signing three years of fixed pricing on a rapidly deflating input is a decision, even when it is made by not thinking about it.

Nobody prices a commodity once and walks away. The unusual thing about this commodity is how quickly walking away stops being an option.

Sources: Chargebee's AI-agent pricing playbook (the seat-pricing "three-body problem" of fast product change, variable per-user consumption, and volatile token cost; outcome-based pricing exemplified by Intercom's Fin billing per resolved issue; action and workflow usage pricing exemplified by N8N's per-workflow batches and Clay's credit burn table; hybrid base-plus-usage models bundling included credits; selection on value attribution, execution autonomy and workload predictability; modeling baseline, spike and supplier-increase cost scenarios before launch; a standing cross-functional pricing committee and continuous recalibration; Cursor's forced move off unlimited plans onto usage caps as the cautionary case); a16z via Chargebee (the cost of an LLM at equivalent performance falling roughly 10x per year).