Full narration of the briefing. Click a section heading to jump the player to that moment.
Most of what we negotiate has a licence and a seat count, so you can at least argue about a number you both agree exists. An Anthropic agreement is not that. The meter is consumption, the unit is a token, and the total depends on choices your engineering teams make every day after you have signed. I am Tom, Claire is with me, and this is Negotiating Anthropic from Redress Compliance.
Six briefings on how to buy this well. And one warning before we start, which we will repeat: this market has almost no public price record, so we will be explicit about what is documented and what is our judgement.
There are three things you might be buying, and confusing them is the most common mistake we see. First, seat based plans for people, where staff use Claude through an interface and you pay per user per month, much like any other software subscription. Second, the API, where your own applications call the models and you pay for what they consume, with no seats involved at all. Third, developer tooling such as Claude Code, which sits between the two.
A finance team that budgets the API like a seat plan will be wrong by an order of magnitude in either direction.
Now the mechanic that decides your bill. Consumption is measured in tokens, roughly speaking fragments of words, and it is counted in both directions: everything you send the model, and everything it sends back. Those two are not priced the same. Output is materially more expensive than input across the published rate cards.
That single asymmetry has real design consequences. An application that sends large documents and asks for short answers behaves very differently on your invoice from one that sends little and generates long prose, even if both look like the same workload on a diagram.
Then model choice, and this is the one to internalise. The family runs from small and fast to large and most capable, and the published price difference between the smallest and largest is roughly an order of magnitude per token. Read that against what a negotiation realistically achieves. No discount you win will move your cost by ten times.
Routing work to the smallest model that is genuinely good enough will. So the highest value conversation is not with the vendor at all. It is with your own engineers, about which workloads actually need the largest model.
There are also three routes to the same models, and the route changes who you negotiate with. You can buy direct from Anthropic. Or you can consume the models through Amazon Bedrock, through Google Cloud, or through Microsoft's platform, because Anthropic publishes availability on those. That third option matters commercially, not just technically.
If you already hold a large cloud commitment, consumption bought through that cloud may draw down a commitment you have already promised to spend. We will come back to this in part five, because it is one of the few genuine pieces of leverage you have.
Two cost reducers are published product features rather than concessions, and you should be using both before you ask anyone for a discount. Prompt caching lets repeated context be reused across calls at a lower rate than sending it fresh every time, which matters enormously for applications with long standing instructions. And batch processing, for work that does not need an immediate answer, is offered at a substantial published discount to the standard rate. Neither of those requires a negotiation.
Both change your baseline, which is precisely why they need to be in place before you size a commitment.
So three things before the next episode. Separate your seat spend from your API spend, because they are different purchases with different economics and probably different budget owners. Get your actual token split, input against output, per workload, because that is the shape of your bill and most organisations have never looked at it. And find out whether you are already paying a cloud commitment that this consumption could draw against.
Next time, Daniel and I turn all of that into a number, which is the hardest part of this whole exercise.
Redress Compliance works on contingency: our fee is 25 percent of what we save you. Nothing saved, nothing paid. Independent, buyer side only, never vendor funded. Want Redress to contact you? Reach out and we respond the same day.
Talk to a Anthropic negotiator