I've been looking for this framework in the public conversation about AI pricing for a year. It doesn't exist. The strategic decision around what to do as inference costs fall under your priced outcomes is one of the most consequential CPO/CFO decisions of the next 24 months, and there's no public playbook.
This essay is the framework. The companion model is in /toolkit/margin-recovery-curve-model.
The short version
Inference costs are dropping ~50% per year (industry-wide). Your outcome pricing was set at one moment in time. As costs fall, the gap between your price and your cost grows. You face a decision: hold the price (capture margin), pass the savings (capture share), or split the difference.
This decision isn't gradual. It comes as a cliff: a competitor undercuts you 12-18 months in, and you have to react. The companies that planned for the cliff handle it cleanly. The companies that didn't lose share or margin instantly.
The framework: model three scenarios over 24 months, track competitor pricing monthly, and decide which scenario you're running deliberately.
Why the cliff is real
The math.
Token costs dropped from ~$10/M tokens (2024) to ~$1/M tokens (2026). Industry trajectory suggests ~$0.50/M tokens by 2027.
If your outcome price is $0.50/resolution and inference costs are 25% of revenue today, in 24 months inference costs will be 5-8% of revenue. GM rises from 60% to 80%+ if you hold price.
That's a 20-point margin lift. Sounds great. Until the competitor with 5% inference cost prices at $0.30/resolution and offers free trial conversion. Now you're sitting at $0.50, your customers are looking at $0.30, and the cliff hits.
The three scenarios
Three sections covering each scenario in detail with the math, the risks, and when each is the right call:
Scenario A: Hold prices
When it works: differentiated product, sticky customers, low competitor threat, high switching costs. Examples: enterprise products with long contracts.
The math: revenue stable, GM rises 20+ points over 24 months. Margin dollars per customer up significantly.
The risk: a 12-month-out competitor undercut that you can't respond to without admitting your old price was inflated.
Scenario B: Pass full savings
When it works: high-volume products, price-sensitive customers, competitive market, growth-stage company.
The math: revenue per outcome drops in proportion to cost. GM stays at trough level. Volume grows because customers can afford more.
The risk: volume doesn't compensate for the price drop. Revenue grows slower than expected.
Scenario C: Pass 50% of savings
The pragmatic middle. Most common choice. Captures half the margin, halves the competitive exposure.
How to plan for the cliff
Four moves:
- Build the 24-month model now (use the toolkit linked above).
- Track three competitor prices monthly. Set a threshold (e.g., "if any direct competitor prices >25% below us") that triggers a price review.
- Decide which scenario you're running. Write it down. Get CFO sign-off.
- Communicate the decision internally. Sales needs to know whether to defend price or follow.
What I'd do if I were starting over
Three things:
- Build the cliff model on day one of pricing strategy, not month 12.
- Set explicit price-review triggers. Don't wait for a customer to churn to wake up to the issue.
- Communicate the strategy to sales early. Sales reps making one-off discounts in the field is the number one way the cliff strategy gets undermined operationally.
Related
- The Pricing Migration Sequence, the strategic context for pricing in transition.
- Margin Watch Agent, the agent that operationalizes Jevons-cliff forecasting.
- Pricing Migration: 18-Month Playbook, the pillar chapter.
Frequently asked
What is the Jevons cliff?+
The strategic decision moment when inference costs have dropped enough that your outcome price is no longer at the sweet spot you priced it at originally. Token costs drop ~50% per year. Your $0.50/resolution price was set when costs were higher. As costs fall, you have to decide: hold price (capture margin), pass savings (capture share), or split the difference.
Why is this a cliff and not a curve?+
Because the decision typically happens discretely, not gradually. Most companies hold their initial outcome price for 12-18 months, then face a competitor who's priced lower. At that moment they have to react: match (lose margin instantly), explain why they're worth more (only works for some products), or accept share loss. The 'cliff' is the moment of forced decision.
What are the three scenarios?+
(1) Hold prices: GM rises as costs fall. Revenue stable. Risk: competitor undercuts. (2) Pass full savings: GM stays at trough level. Customer prices drop. Volume grows. Risk: revenue grows slower than expected. (3) Pass 50% of savings: balanced. Most common choice.
How do I model the cliff for my product?+
Three inputs. Current GM per outcome. Inference cost trajectory (industry-wide ~50% per year, your routing improvements may add another 10-25%). Competitor pricing trajectory. Run the three scenarios over 24 months. The model surfaces when the cliff hits and what it costs.
What if I'm wrong about competitor pricing?+
You will be. The fix is to track actual competitor prices monthly and update your model. The Jevons cliff is dynamic; 'set and forget' is the failure mode. Most teams price once at launch and don't revisit until a customer churns to a competitor 8 months later. By then the cliff already happened.