Agency pricing and volume discounts: how to negotiate, and how far it goes

First, understand where the discount comes from

A relay site's quote is upstream cost plus markup. So the room it has to discount you depends on the discount it gets on supply.

• If it resells official quota, its upstream cost is the official price times its own purchasing discount. Its markup room is the gap between official price and its cost. Whatever it gives you must be smaller than that gap.
• If it runs mixed sourcing, its upstream cost is lower and more volatile, so there is more room to discount — but also a weaker stability promise. On these sites, cheap and stable are negatively correlated.
• If it runs reverse-engineered subscriptions, the cost structure is entirely different, amortised over subscription fees, and the discount room is large. But these sites should not be carrying a production workload in the first place.

So the first job is not to negotiate, it is to work out which category the site belongs to — see our article on official, mixed and reverse sourcing. Negotiating a long-term discount with a reverse-sourced site means prepaying more money for a service that may fail entirely.

Three pieces of leverage you hold

One: a stable monthly volume commitment

The most effective lever. What a relay site fears most is traffic volatility, because it has to provision capacity and may hit upstream rate limits. Stable volume has real value to it.

How to use it: show screenshots of your actual usage over the past two months, from your own statistics. That is far more convincing than a verbal promise.

Two: prepayment

Prepayment improves its cash flow, so it usually buys an extra discount.

But this is the highest-risk lever of the three. See our article on the five terms to check before topping up. The precondition for prepaying is that you can accept the money possibly not coming back.

A middle path: accept prepayment but ask for the discount to be settled quarterly, rather than locking a discount into a long contract in one go.

Three: a long-term intent, paired with exit terms

If you can commit to a volume range over 6 to 12 months while requiring monthly adjustability, that is reasonable for both sides.

One caution: do not sign a contract with "no volume adjustment, no exit" just to get a discount. Your volume forecast is very likely wrong, and once locked in you can only grit your teeth.

How far it goes

This is the practical question. Based on public procurement information and common industry practice, here is a rough reference range. It is for judging reasonableness, not a promise.

The criterion: if the discount offered is far above the usual range for that volume, suspect its cost structure first rather than celebrating. Extremely low prices usually mean one of two things:

• an unstable upstream source, reverse-engineered or a channel about to expire;
• or a "discount" that applies only to one specific model, which in practice gets routed to a cheap backend.

Four kinds of quote to be wary of

One: a multiplier with no absolute price

"0.6x for you" sounds better than "20% off", but the multiplier's baseline is set by the site itself. Always convert to "what do I actually pay per million tokens" before comparing.

Two: a discount tied to one model

"Sonnet at 40% off", but the model you need is Opus. That kind of quote looks generous and solves nothing.

Ask for: a separate quote for each model you will actually use.

Three: discounts that step with top-up amount

"Top up 5,000 for 20% off, 20,000 for 40% off." This is fundamentally using the discount to induce you to prepay more. Back to the first principle: the risk of prepayment sits with you.

Four: "unlimited" plans

Any form of monthly unlimited plan deserves suspicion in a relay context, because its own upstream is not unlimited. If it dares to sell unlimited, it is betting you will not use that much, or it plans to throttle or degrade once usage exceeds expectations.

Four things to confirm in writing after negotiating

Verbal promises do not count. At minimum confirm:

1. The list of models the discount applies to — which ones are discounted and which are not.
2. The discount's validity period and adjustment mechanism — the rule by which it moves when upstream prices move.
3. Quota and concurrency commitments — not "unlimited", but specific TPM and RPM figures.
4. Balance refund conditions — what happens to the remaining money if you stop using it.

Item 4 is the most easily skipped, and the most likely to become a dispute at exit.

An easier alternative

If your volume has not yet reached the level where negotiation pays, spend the time on two other things instead:

1. Understand your usage structure. Which calls could go to a smaller model, which could use caching. Those two optimisations usually save far more than any discount.
2. Keep a second supplier configured. Not in use, but configured and ready to switch. That is both availability insurance and negotiating leverage — being able to leave at any time is worth more than any script.

The cost walkthrough in our article on going from official price to your bill will help you work out which calls are most worth optimising. Optimise the structure first, then negotiate the price. Reverse the order and you will spend your time on a 5% discount while missing a 50% structural saving.