Full narration of the briefing. Click a section heading to jump the player to that moment.
In a licence agreement the negotiation decides what you pay, and then you pay it. In a consumption agreement the negotiation decides the rate, and a year of engineering decisions decides the invoice. That is why this last episode matters as much as the other five together. I am Tom, Claire is with me, and this is the discipline that determines what your term actually costs, as opposed to what your contract said it might.
Start with attribution, because everything else depends on it. Every call should carry enough tagging that you can say which application, which team and which feature caused it. Then a monthly view that shows consumption by workload, by model tier and by input against output. If you cannot produce that, you cannot tell an efficiency saving from a drop in usage, and you cannot answer the question your finance team will certainly ask, which is why the bill moved.
This is a week of engineering work that pays for itself in the first quarter.
Then routing, which as we said in part one is worth more than any discount. Review workloads on a schedule and ask whether each one still needs the model it is using. Two things change the answer over time: your prompts get better, and new models arrive that do more for less. A workload that genuinely needed the largest model last year may not this year.
The organisations that control this cost well treat routing as a standing review rather than a one off decision made when the application was built.
Third, keep caching and batching current, because both decay. Caching gets applied to the prompts that existed when someone set it up, and then new prompts arrive without it. Batch eligible work gets built as real time by default because that was easier at the time. So audit both on a cadence: what fraction of your eligible context is actually being cached, and what fraction of your deferrable work is actually running as batch.
Those two percentages are the cleanest efficiency metrics you have in this category.
Fourth, manage the commitment as a running position rather than an annual surprise. Track actual burn against the commitment curve every month, and project the year end position from that. If you are running under, you have time to either accelerate genuine adoption or open a conversation about the ramp, and both of those are only available early. If you are running over, you want to know before an overage rate applies rather than afterwards.
Nobody has ever benefited from discovering either of those in month eleven.
Fifth, put ordinary engineering guardrails in place, because consumption pricing punishes accidents in a way licence pricing does not. Sensible limits per application, alerting when a workload moves outside its normal range, and a review before any new integration goes to production. We have seen a single misconfigured retry loop produce a month of unplanned spend. None of that is exotic and all of it is cheaper to add now than to explain later.
And the last point, which ties the series together. Everything you instrument during this term becomes your evidence for the next negotiation. Clean attribution, a demonstrated efficiency record, and a defensible forecast are worth more at renewal than any argument you can construct in the final month. So the work in this episode is not administration, it is the position you will negotiate from next time.
That is the series. If you are sizing a commitment or reading an agreement now, talk to us before you sign it.
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