Should You Pass an AI Price Cut to Your Client?

On 22 September 2026 both OpenAI and Anthropic cut their API prices, in OpenAI's case by half. If you are running a client project on either, your cost base moved overnight and your invoice did not. Whether you should pass that saving to your client is not a matter of conscience. It is a matter of which contract you signed, and the three common shapes give three different answers.

Start with what the contract says the client is buying

The question is never "is it fair to keep the saving." It is "what did the client agree to pay for." Those come apart quickly.

Most disputes come from people treating the third row as if it were the first.

Fixed fee for an outcome

You quoted £12,000 to build a document processing pipeline. The client agreed to £12,000 for a pipeline that works. Your model costs are an input to your estimate, not a term of the deal. If the provider had doubled prices in month two you would have absorbed it, and nobody would have suggested the client top you up. Symmetry cuts both ways: you carried the price risk, so you keep the price windfall.

That is the correct answer commercially and it is still worth telling them. More on that below.

Cost-plus and passthrough

If your agreement says you bill API usage at cost, or at cost plus a stated percentage, then the cost is a defined term and it just changed. Continuing to invoice at the old rate is not a pricing decision, it is an incorrect invoice. Reprice from the date the new rates took effect, and if you have already invoiced across that boundary, issue the credit without waiting to be asked. The mechanics of tracking this cleanly are in how to bill clients for AI API usage.

Retainer with an allowance

This is the ambiguous one, and the ambiguity is usually your own doing. If you sold "£2,000 a month including up to 50,000 documents processed," you sold capacity and the price of that capacity is yours to set. If you sold "£2,000 a month, which covers roughly £600 of model usage plus my time," you itemised your costs to justify the number, and the client will reasonably expect the £600 line to fall. You cannot use cost transparency to win the deal and then treat costs as none of their business when they move in your favour.

If you are unsure which of the two you did, go and read the proposal you sent. The answer is in there, and it will be more specific than your memory of it.

The case where keeping it quietly costs you more

There is one scenario where the commercially correct answer still loses you money: the client finds out from somewhere else.

A 50% API price cut announced by OpenAI is not obscure industry gossip. It was covered widely, and a client who is spending meaningfully on an AI project is exactly the kind of person who reads that coverage. If they learn about a halving of your input costs from a news article three months after it happened, while you were invoicing unchanged, the fixed-fee argument becomes very hard to make in that conversation. It is correct and it will sound like an excuse.

The asymmetry is brutal. On a fixed-fee project the saving might be a few hundred pounds a month. The renewal is worth many times that. Trading the second for the first is a bad deal even when you are entitled to the first.

What to actually do

1. Work out the number before you decide anything. Pull last month's usage, reprice it at the new rates, and get a real figure. It is frequently smaller than it feels, because rate cuts apply to tokens and your bill includes fixed costs, your time, and paths that did not change. Sometimes the whole question is worth £40 a month and you can stop thinking about it.

2. Tell them either way, in one paragraph. "You may have seen that model prices dropped this week. On your project that works out to about £180 a month. Under our fixed-fee agreement that sits with us, and it is also what would have happened to me if prices had gone the other way. Flagging it so you hear it from me." That paragraph protects the renewal and costs you nothing.

3. Consider spending it rather than keeping it. On a fixed fee you can convert the saving into something visible: move a path to the better model, raise a quality threshold, add the evaluation suite you skipped. The client gets more, you keep the margin structurally, and at renewal you have a story about what you did with it.

4. Fix the contract for next time. Add one clause specifying what happens when provider pricing moves more than some threshold in either direction. Both directions. A clause that only protects you when costs rise is one the client will notice and resent later. The companion problem is covered in what to do when your AI provider raises prices.

Three things that change the answer

The three-row table is the default. Four situations override it.

You quoted the project during the promotional window. If your estimate was built on a rate the vendor had flagged as temporary, and that rate has now become permanent or fallen further, you priced in a risk that did not materialise. That is closer to a contingency you did not spend than to a windfall, and clients who understand the distinction will notice.

The client supplies the API key. If they hold the account and you build against it, the saving was never yours in the first place. It has already reached them automatically. The only thing to do is tell them it happened, because they may not have noticed, and it is a free piece of good news with your name on it.

There is a volume commitment in play. If either side committed to a spend level to unlock a rate, a price cut can put you under the threshold and trigger a worse rate. Check the commitment terms before you celebrate. This is the one case where a price cut can cost money.

You are mid-renegotiation. If you are already in a conversation about scope or fees, the saving is part of that conversation whether or not you raise it. Raising it yourself is worth more than having it raised at you.

Do not let the saving hide a worse problem

A price cut can mask cost growth. If your usage rose 30% last quarter because an agent loop is resending context it does not need, a 50% rate cut turns a visible problem into an invisible one, and the underlying inefficiency stays in the system until usage grows enough to surface it again. Before you decide who gets the money, spend an hour on reducing your actual API costs. Structural savings survive the next price change; rate savings do not.

Two labs cut prices on the same day, which suggests the direction of travel for the next year is downward and irregular. Build the answer into how you quote rather than deciding it fresh each time, which is the point of thinking about engagement pricing structurally and of the wider material on making money with AI work.