A 60% provider price cut changes what you build, not just the margin

The provider announcement landed at 16:40 on a Wednesday and the engineering channel had reached a conclusion by 16:52: input and output rates down 60%, effective on the next billing cycle, no action required. Someone posted the current month's model spend with a strikethrough and the new number beside it. Someone else asked whether this meant the pricing work could be shelved.

The saving is real and worth about $12,000 a month on the seat product, and taking it is the least interesting thing available. A cut of that size moves three things that have nothing to do with this quarter's margin: which features are worth building, which cost lines now dominate, and what your competitors can now afford, since they received the identical discount on the identical day.

Recomputing the stack, including what does not move#

Fleet arithmetic first, because the number people will quote is usually wrong in a predictable direction. Revenue is 4,180 seats at $29, or $121,220 a month. Model spend was $20,273 and becomes $8,109. Non-model cost of goods — support contacts, escalation review, retrieval infrastructure, allocated platform — is $4.30 per seat, or $17,974, and none of it moves.

Ten points of margin from a 60% input cut, because a little over half the cost base was exposed to it. The same calculation on the agent product from chapters 3 and 4 takes cost per completed session from $0.875 to $0.669 and margin from 23.7% to 41.7%, a bigger swing because that product is more model-heavy, and still not the doubling that the headline number suggests to anyone doing the arithmetic in their head.

The last row is the one to carry forward. Model spend is now under a third of cost of goods on the seat product, which means the engineering effort that has been going into token reduction has lost about half its leverage overnight. A 20% improvement in cache hit rate used to be worth $2,100 a month and is now worth $840. A one-percentage-point reduction in escalation rate — from 4.1% to 3.1% — is worth $2,080 a month and was worth exactly the same yesterday. Whichever line is largest is where the next quarter of cost work belongs, and the price cut has just reordered the list.

Two contractual details are worth checking in the same week. If you hold a committed-spend agreement negotiated at the old rates, your commitment is denominated in dollars and your consumption just fell by 60%, which means you are on course to under-consume and pay the effective premium chapter 4 described; a commitment of $200,000 a month against $80,000 of post-cut usage is an effective rate 2.5× list. Most providers will restructure rather than watch that happen, and the conversation is much easier before the shortfall appears than after. If instead the commitment is denominated in tokens or throughput, check what the cut did to the price of that unit, because reserved capacity does not always move with list rates.

The frontier moves before the margin does#

Every product has a column of things that were evaluated, priced, and declined because the cost per unit did not clear the value per unit. That column is where a 60% cut does its real work, because each item in it has a threshold and some of those thresholds have now been crossed.

The verifier is the instructive row. It was declined eight months earlier on arithmetic that was correct at the time: $0.060 of extra inference to prevent $0.035 of reviewer time is a losing trade, and the team was right to say no. At the new rates the same feature saves $0.011 per session, which across 190,000 sessions is $2,090 a month, and it also removes about 1,300 escalations from a review team that is the constraint on how fast the product can onboard new customers. Nothing about the feature changed. The threshold moved under it.

This is why the response to a price cut with the highest return is not a finance exercise. It is going back through the rejected column with the new numbers, which takes an afternoon if you kept the estimates and a week if you did not, and is the single strongest argument for writing down the cost basis of a rejection at the moment you make it. A rejection recorded as "too expensive" is unusable a year later; one recorded as "$0.060 per session against $0.035 of avoided review, revisit below $0.032" re-evaluates itself.

The third row is a different kind of move. Where a smaller or distilled model — a compact model trained to reproduce a larger one's outputs at a fraction of the price — was chosen to fit a budget, a cut re-opens the question of which model the workload should have been running on, and the answer is often that the compromise is no longer necessary for the hard slice of traffic. Routing eighteen percent of requests to a stronger model at $0.033 a session, with escalation on that slice falling from 11% to 6%, is cheaper in total than the configuration it replaces.

What the cut does not buy#

It buys no competitive advantage, because it was not a discount, it was a price change, and every company building on the same provider woke up with the same cost structure. Historically these cuts propagate to end-customer pricing over two to four quarters, through a competitor undercutting on price or through buyers arriving at renewal having read the same announcement and asking why their bill has not moved. Treating the saving as permanent margin is a forecast that the market will not notice, and the more defensible use of it is to convert it into something a competitor cannot get from a price list: lower escalation rates, better retrieval, a verifier that makes the product measurably more reliable.

It also does not fix the structural problem in flat pricing. The $340 maximum seat becomes a $136 seat, and break-even on a $29 plan sits at $24.70 of model spend, so that seat is still 5.5× underwater. The distribution scaled; its shape did not change, and the mechanism generating the tail — automation with no ceiling — is untouched. The allowance structure from chapter 7 is still the right structure; only its parameters need re-deriving, since 150 credits sized against $0.19 sessions can be more generous against $0.076 sessions, and a more generous allowance is a genuinely good thing to give customers at renewal.

There is a second-order effect that runs the other way and is usually omitted from the celebration. Cheaper inference makes automation more attractive to your customers as well, and to your own product managers, so consumption tends to rise after a price cut rather than staying flat. A 60% reduction in unit price against a 2.5× increase in units consumed is a 0% change in spend, and that is not a hypothetical: the features in the table above are themselves consumption increases you are choosing to make. Model the volume response explicitly rather than assuming the saving banks itself.

The volume response is worth modelling with the same discipline as the cost side, because it is estimable rather than mysterious. Three of the four features in the table above are consumption increases you control, and they can be priced before they ship: the verifier adds a call to every session, re-ranking adds one, the nightly summaries add 4,180 a day. Together they raise sessions-with-inference by roughly 2.3× while unit price falls 60%, which lands post-cut model spend at about $18,600 rather than $8,109 — most of the saving spent deliberately, on things with measured returns. That is a defensible plan and it is a very different sentence from "margins improved ten points", and the difference is entirely in whether anyone wrote the second-order effects down.

Finally, price cuts frequently arrive attached to a new model version, and adopting it is not free. An evaluation run across your regression set — the fixed collection of cases you re-run on every change to catch behaviour that got worse — a shadow deployment, prompt adjustments where behaviour differs, and the risk that a workflow tuned over six months regresses in a way your evals do not cover — that is typically two to four engineer-weeks for a mature product, and it belongs in the decision alongside the saving.

What to do with the model you have built#

The eight chapters of this topic assemble into one artefact: a spreadsheet or a query that starts from component token counts, prices a session including the ones that failed, adds the human and platform costs that never appear on an invoice, allocates shared pools by a driver, and reports the result per tenant with the distribution rather than the mean. Its value is that it can be re-run, which is exactly what a Wednesday afternoon price cut demands.

Re-run it now with the new rates. Re-run the sensitivity from chapter 5, because the parameters moved and the break-even did not. Re-rank tenants by cost per contract dollar, since the seven accounts that were above threshold may now be three, and the two that remain are the ones whose usage was never really about price. Re-derive the allowance. Then take the rejected column and price it again, because that is the part of the exercise that produces something a customer can see.