The Grandfather Problem: How to Reprice Agentic AI When Model Costs Keep Falling

Table of Contents

• What the Grandfather Problem Actually Is
• Why Falling Model Costs Break Your Pricing Faster Than You Think
• The Two Directions a Repricing Can Go
• The Four Levers You Actually Have
• Lever 1: Hold the Price, Pocket the Margin
• Lever 2: Pass the Savings Through
• Lever 3: Re-Anchor on Value, Not Cost
• Lever 4: Migrate the Plan, Not the Price
• How Pricing Architecture Decides Your Fate in Advance
• Running an Actual Repricing: A Sequenced Playbook
• Insights Most People Overlook
• Frequently Asked Questions
• Conclusion
• References

What the Grandfather Problem Actually Is

In classic SaaS, "grandfathering" means letting old customers keep an old, usually cheaper, price when you raise rates. It's a goodwill move and a churn shield. The grandfather problem in agentic AI-as-a-service is a different animal, and most teams don't see it coming because they're braced for the SaaS version.

Here's the setup. You sell an agent, say, a support-resolution agent priced at $2 per resolved ticket. When you set that price, your underlying cost per resolution was about $0.70: model inference, orchestration, retrieval, retries, the works. A 65% gross margin felt healthy. Eighteen months later, the frontier model you depend on has gone through two major price cuts, you've added a routing layer that sends easy tickets to a cheaper model, and your real cost per resolution is now $0.22. That same $2 price now carries an 89% margin.

That sounds like a champagne problem until you realize what it does to your competitive position. A new entrant can price at $1.20 and still clear better margins than you did at launch. Your "grandfathered" enterprise accounts, the ones on annual contracts you were proud to land, are now sitting on prices that look indefensible the moment a procurement team benchmarks them. The asset (a fat margin) and the liability (a price that invites churn and undercutting) are the same number. That's the grandfather problem: falling costs silently convert your pricing from competitive to exposed, and you have to decide what to do about it before a competitor or a CFO decides for you.

Why Falling Model Costs Break Your Pricing Faster Than You Think

Software pricing has always assumed roughly stable unit costs. The marginal cost of serving one more SaaS seat is near zero and stays near zero, so you price on value and forget about cost. Agentic products broke that assumption in both directions: unit costs are real (every task burns tokens), and they're falling fast and unpredictably.

The pace is the part people underestimate. Token prices for frontier-class models have dropped by more than an order of magnitude over a couple of years, a trend Andreessen Horowitz has tracked as roughly a 10x annual decline in the cost to reach a given level of capability. Stack three forces on top of raw price cuts and the deflation compounds: smaller open-weight models catching up to last year's frontier, your own model-routing logic getting smarter about when to use the cheap option, and prompt-caching plus batching shaving real money off every call. A team that does nothing but ship the same product for a year can watch its cost-per-task fall by half through no deliberate effort.

The result is that your pricing has a half-life now. A price you set on sound cost-plus logic in Q1 can be 30 points of margin too generous by Q4. SaaS founders are used to pricing being a thing you revisit annually, almost ceremonially. In GaaS, the ground under the price moves on a quarterly cadence, and the customers who locked in early move the slowest. This is why repricing isn't a one-time correction, it's a discipline, closely tied to the broader question of passing volatile inference costs through to customers without torching trust.

The Two Directions a Repricing Can Go

People hear "repricing" and assume "raise prices." With the grandfather problem, the more interesting and more common move is the opposite, and getting the direction right matters more than the mechanics.

Repricing down is what falling costs invite. You cut the customer's effective price, sacrificing some of your windfall margin to stay ahead of competitors and to lock in loyalty. This is defensive and proactive: you'd rather hand back ten points of margin on your own terms than lose the account to a cheaper entrant or a renewal-time benchmark fight.

Repricing up still happens in GaaS, but for a narrower reason: usage growth. If your customer's agent usage has tripled and they're on a fixed annual fee that no longer covers the load, you reprice up at renewal even as your per-task costs fall. The two effects can cancel or compound depending on the account. This is the annual-contract problem when usage is unpredictable showing up in a new costume.

The decision that actually matters isn't up versus down, it's who moves and who doesn't. You almost never want a blanket reprice. You want a segmented one: hold the line on price-insensitive accounts that aren't benchmarking, cut proactively for the sophisticated accounts most likely to notice and shop around, and restructure the handful where usage has outgrown the contract. Treating repricing as one global lever is the first mistake; it's really a portfolio of per-segment decisions.

The Four Levers You Actually Have

Lever 1: Hold the Price, Pocket the Margin

The simplest move is to do nothing and enjoy the expanding margin. For a real slice of your base, this is correct. Customers who don't benchmark, who value the outcome far above the price, and who'd incur real switching costs to leave are not worth proactively discounting. Margin you give them back is margin you simply lit on fire.

The risk is asymmetric and worth naming. Hold too long across too many accounts and you're building a business on prices that one well-prepared procurement deck can detonate at renewal. Holding is a fine default for the bottom-left of your account matrix, low sophistication, high stickiness, and a dangerous one for the top accounts. The discipline is knowing which is which, not picking a single policy for everyone.

Lever 2: Pass the Savings Through

Here you cut the price to reflect lower costs, either proactively or at renewal. The upside is goodwill and competitive insulation: a customer who sees their per-task price drop without asking is a customer who isn't taking discovery calls from your competitor.

The trap is the ratchet. Once you've trained an account that prices fall automatically as your costs fall, you've made your margin a function of model-provider roadmaps rather than the value you deliver, and you've handed the customer a permanent expectation that only moves one direction. Pass-through works best as a deliberate, communicated gesture ("we cut your rate 15% because our costs came down") rather than a silent, formula-driven slide that becomes the baseline expectation forever. If you go this route, cap it. Decide in advance how much of any future cost decline you'll share, and say so.

Lever 3: Re-Anchor on Value, Not Cost

The strongest lever is to stop letting cost set the conversation at all. If your agent resolves a support ticket a human would've spent 12 minutes and $9 of loaded labor on, the defensible price has almost nothing to do with your $0.22 inference cost, it's a fraction of the $9 you displaced. Re-anchoring means migrating the customer's mental model from cost-plus to value-based, which is the entire premise behind pricing when the value delivered is a replaced employee.

This is the move that makes the grandfather problem disappear, because falling inference costs stop being relevant to the price. The catch: you can only re-anchor credibly if you can measure and prove the value, which is its own discipline, the territory of outcome-based pricing and who gets to define and audit the outcome. Vendors who never built that measurement muscle are stuck defending a cost-plus number that erodes every quarter. The companies that escape the grandfather trap usually did the value-instrumentation work early, before they needed it.

Lever 4: Migrate the Plan, Not the Price

Often the cleanest path is to leave the legacy price technically intact while moving the customer onto a new structure, a credit pool, a hybrid base-plus-usage model, or a tier with a different metric. You're not raising or cutting their price so much as changing the unit the price attaches to, which sidesteps the loss-aversion that any explicit increase triggers.

This is where pricing architecture pays off. Moving a customer from per-task to a prepaid credit pool or a hybrid subscription-plus-usage model lets you reset the economics under cover of "we've upgraded your plan" rather than "we're charging you more." Migration is the diplomat's lever: it does the work of a reprice while looking like an upgrade. Done well, the customer experiences a better-fitting plan; done clumsily, it reads as repricing in a trench coat, and the trust hit is worse than an honest increase would have been.

How Pricing Architecture Decides Your Fate in Advance

Most of your repricing freedom is determined before you ever face the problem, by how you structured pricing on day one. This is the part founders wish they'd known.

If you priced per-outcome from the start ($2 per resolution), falling costs just widen your margin and you have maximum freedom: hold, trim, or re-anchor at will, because the customer was never paying for your tokens. If you priced transparent cost-plus (tokens used × a published markup), you've welded your revenue to model-provider price cuts, every decline mechanically drops your revenue, and customers will notice and expect it. That's the transparency trap of showing token counts biting back. And if you sold a flat annual fee, you're insulated from cost deflation but fully exposed to usage growth, which is why some GaaS vendors are quietly returning to flat pricing precisely to escape the cost-tracking treadmill.

The strategic takeaway is uncomfortable but clear: the more your pricing metric is decoupled from inference cost, the less the grandfather problem can hurt you. Outcome and value metrics are deflation-proof; token-metered and cost-plus metrics are deflation-exposed. If you're early enough to still be choosing, choose the metric that won't force you into an awkward reprice every time a model provider issues a press release. The GaaS pricing taxonomy of per-task, per-outcome, and per-seat isn't just a billing detail, it's the thing that decides how often you'll be sitting in this exact meeting.

Running an Actual Repricing: A Sequenced Playbook

When you do decide to move, sequence matters more than the size of the change. A rushed, undifferentiated reprice churns customers who'd have happily stayed.

Segment first. Build a simple two-axis grid: price sensitivity / benchmarking sophistication on one axis, current margin on the other. The fat-margin, high-sophistication accounts are your priority, they're the ones most likely to notice and shop. The low-sophistication, reasonable-margin accounts can usually be left alone.

Decide direction per segment, not globally. Some accounts get a proactive cut; some get a structure migration; most get nothing. A blanket reprice is almost always wrong.

Lead with value, never with cost. The fastest way to lose a repricing conversation is to let "your costs went down" become the customer's opening line. Re-anchor on outcomes delivered before anyone reaches for an inference-cost calculator. McKinsey's work on the economic potential of generative AI is useful ammunition here: the value frame, measured in displaced labor and accelerated work, dwarfs the token frame by orders of magnitude.

Time it to a renewal or a feature release. Repricing inside a contract term reads as a grab. Repricing at renewal, or alongside a genuine capability upgrade, reads as a recalibration. Always attach the change to a moment that gives the customer a reason beyond "we wanted more."

Pilot, then roll. Test the new structure on a cohort, measure churn and expansion, and only then generalize. This connects directly to what top startups are A/B testing in their pricing experiments, repricing is just an experiment with existing customers, and it deserves the same rigor.

Insights Most People Overlook

The grandfather problem is a leading indicator of a commoditization threat, not just a margin question. If your costs have fallen enough that holding price feels uncomfortable, a competitor can almost certainly enter below you and still profit. The expanding margin you're enjoying is the same signal that tells you the moat is thinning. Read it as a warning, not a windfall.

Silent pass-through is worse than no pass-through. Quietly lowering a customer's price as your costs fall feels generous, but it trains the account to expect automatic, perpetual decreases and gets you zero goodwill credit because they never knew it happened. If you're going to give margin back, make them feel it, an announced 15% cut buys loyalty that a silent 25% slide does not.

Your most loyal early customers are your most dangerous repricing risk, not your safest. Teams assume long-tenured accounts will tolerate a reprice out of relationship equity. The opposite is often true: they remember the original price, they've watched the model-cost headlines, and they feel entitled to the deflation. Loyalty raises their expectation of fairness, not their tolerance for paying more than a newcomer.

Outcome pricing doesn't eliminate the grandfather problem, it relocates it to your outcome definition. If your "resolution" or "qualified lead" definition was calibrated when models were weaker, falling costs and rising capability mean each unit now takes less work to deliver. The savvy move isn't repricing the unit; it's revisiting what counts as a billable outcome, which ties straight into pricing for partial completion and graceful degradation.

Model-cost deflation quietly subsidizes your reliability spend. As inference gets cheaper, you can afford more retries, more verification passes, and more redundant model calls per task at the same total cost, buying back margin as reliability instead of as cash. The vendors who reinvest deflation into agent reliability often out-retain the ones who bank it, because reliability is what actually drives expansion.

Frequently Asked Questions

Should I ever proactively cut a price the customer hasn't complained about? Yes, but only for the segment most likely to benchmark and churn, and only when announced clearly enough to earn goodwill. For sticky, low-sophistication accounts, a proactive cut is usually just donated margin.

How do I reprice without inviting a renegotiation of the entire contract? Attach the change to a renewal or a capability release, and migrate the plan structure rather than editing the headline number. Changing the metric ("you're now on a credit pool") triggers far less loss aversion than changing the price on the existing metric.

What if my biggest customer is on the most generous legacy price and explicitly knows model costs have dropped? Re-anchor on value before they re-anchor on cost. Quantify the labor or revenue outcome you deliver, propose a structure migration that improves their fit, and avoid a line-item argument about your inference bill, which you will lose.

Does outcome-based pricing make me immune to all this? It makes you immune to cost-deflation pressure on the price, but it shifts scrutiny onto your outcome definition and your ability to measure it. The grandfather problem moves from "your price is too high for your costs" to "your outcome definition is stale for today's capability."

How often should I revisit agent pricing given how fast costs move? Treat it as a standing quarterly review of cost-per-task versus price, with actual customer-facing changes timed to renewals. The internal monitoring is continuous; the external repricing is deliberate and infrequent.

Is cost-plus pricing just a mistake for agent products? Not always, it's transparent and easy to sell early, but it welds your revenue to model-provider roadmaps and maximizes your exposure to the grandfather problem. If you use it, plan your escape to a value or outcome metric before deflation forces an awkward conversation.

Conclusion

The grandfather problem is what happens when an old assumption, stable unit costs, collides with a new reality where inference gets dramatically cheaper every few months. Falling model costs don't just pad your margins; they quietly convert yesterday's smart price into tomorrow's competitive liability, and they do it fastest on the loyal, locked-in accounts you were proudest to win. The four levers, hold, pass through, re-anchor, migrate, are each right for a different segment, and the worst move is treating repricing as one global switch instead of a portfolio of per-account decisions.

The deeper lesson sits upstream of any individual reprice: the pricing metric you chose on day one largely determines how much pain this causes you. Decouple price from token cost, through outcome framing, value anchoring, or smart structure, and model-cost deflation becomes a margin gift you control rather than a liability that controls you. In the broader GaaS pricing and monetization landscape, repricing as costs fall isn't an edge case. It's the recurring tax on every vendor who priced against a number that won't sit still, and the discipline of handling it well is fast becoming a core competency of every serious agent business.

References

• LLMflation: LLM Inference Cost Trends, Andreessen Horowitz
• The Economic Potential of Generative AI: The Next Productivity Frontier, McKinsey & Company