What to Do When Your AI Provider Raises Prices

Start with arithmetic, not with your pricing page. Work out your cost per active user per month at the new rate, compare it to what that user pays you, and you will know within ten minutes whether this is an annoyance or an existential problem. Most teams skip that step, go straight to panic or to a price rise, and end up making a large decision from a feeling.

This is topical because it is about to happen on a schedule. Google's Gemini 3.8 Flash, announced on 2 September, carries an introductory rate of $0.75 per million input tokens that becomes $1.50 after 31 December 2026. That is a doubling, with four months of notice. Anyone building on it now has an unusually clear deadline.

The break-even worksheet

Four numbers. Get them before you do anything else.

1. Tokens per action. Instrument one representative action end to end, input and output separately, including retries and system prompt. Estimates here are wrong by two to five times in both directions.

2. Actions per active user per month. Take the median, not the mean. A handful of heavy users will distort a mean and make you price for a customer you mostly do not have.

3. Cost per user at the new rate. Multiply out. Keep input and output separate, because output usually costs four to five times more per token and is the line you can actually shrink.

4. Gross margin per user. Revenue per user minus that cost, minus everything else variable. If this is still comfortable, stop here and get on with your week.

A worked shape, using round numbers so the mechanics are visible:

A jump from 13% to 26% of revenue is serious but not fatal on a $29 plan. The same doubling on a $9 plan takes you from 42% to 83% and the product does not work any more. Which of those you are in is the only thing that decides what you do next.

The four levers, in order

1. Cut tokens before you cut anything else

This is free and it is almost always available. Typical wins on a first pass: a system prompt that has grown by accretion and can be halved, retrieved context that sends ten chunks where four would do, conversation history replayed in full when a summary would serve, and retries that re-send everything rather than the failed portion.

Cutting your token count by 40% fully absorbs a 66% price rise. Teams that have never looked usually find 30% to 50% in an afternoon.

2. Fix your cache hit rate

Cached input costs a fraction of fresh input, and providers keep widening that gap. Anthropic cut cache reads to $0.25 per million with Claude Fable 5.1 on 1 September, a 75% reduction. Caching rewards stable prefixes, so put the invariant parts of your prompt first and the variable parts last. Reordering a prompt is a smaller change than changing supplier.

3. Route by task, not by habit

Most products use one model for everything because that is how the prototype worked. Classification, extraction and routing rarely need your flagship. Moving the 70% of calls that are mechanical onto a cheaper model, while keeping the 30% that users judge you on where they are, usually beats every other lever combined. How to reduce AI API costs goes through the mechanics.

4. Change your own prices, last

It is the lever with the highest cost of being wrong, which is why it goes last. If you get here, the specifics of doing it without churning your base are in how to raise prices on an AI product. Note the asymmetry with the opposite case: when costs fall you have a genuine choice about passing it on, which is its own decision. When costs rise you usually do not.

Do not switch provider first

The instinct is to move. Before you do, price the move honestly: re-running your evals, rewriting prompts that were tuned to one model's quirks, retesting tool calling, and the risk that the new provider raises its own prices in six months. A 30% saving that costs three engineer-weeks and introduces a quality regression is not a saving.

Switch when the gap is large and durable, not when it is 15% and recent. Headline per-token rates are the least reliable part of the comparison, because they say nothing about how many tokens your workload actually needs on each model.

Build for the next one

A price rise is a stress test of a decision you made earlier: whether your product is coupled to one model. Two habits make the next one boring.

• Keep a per-feature cost dashboard, not a single monthly total. When a provider moves, you want to know in an hour which features are affected and by how much.

• Keep one alternative model working in your evals even if you never use it. It is a small standing cost that turns a migration into a switch.

The related failure to plan for is availability rather than price, covered in what to do when your AI provider goes down. The wider commercial picture is in our AI monetization pillar.

Frequently asked questions

Should I tell customers my costs went up?

Only if you are changing what they pay. Explaining supplier costs as justification invites the argument that you should have hedged them, and customers price on value rather than on your input costs.

How much notice do providers usually give?

Very little historically, which is what makes a published expiry date like 31 December unusual and worth planning around.

Is it worth pre-buying capacity at the old rate?

Only if the provider genuinely offers committed-use pricing and your volume forecast is solid. Committing to volume you do not use is a more expensive mistake than the price rise.

What if the rise makes one feature unprofitable?

Price that feature separately or meter it, rather than raising the base price for everyone. Usage-based add-ons exist for exactly this shape of problem.