Skip to content
ABTO Guide

How-to guides

Respond to a cost spike

Narrow the cause of a sudden cost increase from the feature to the variant to a single call.

It starts with one invoice that is 40% higher than last month. Staring at the total only gets you “let’s all spend less”, so narrow the spike down three times instead.

Get the cost ranking by feature from Insights (how to ask is covered in See spend by feature). What matters here is the follow-up “when did it start climbing?”, which pins down the date.

A step up from a specific date points at a deploy or change made that day. A steady climb points at traffic growth or a shift in usage patterns. If outcomes grew alongside, this is growth; cost rising on its own is what needs action.

In Compare, put the feature’s variants side by side and read the cost per call. Each variant is a different model and prompt, so the table shows which combination is driving the cost.

If the feature runs a single variant, there is nothing to compare, so go straight on.

In Requests, filtering by the model the previous step pointed to leaves only that variant’s calls. Opening one call’s detail shows the token and cost breakdown, which tells you whether the input or the output grew.

SymptomCommon causeAction
Prompt tokens spikingConversation history or context stuffed in wholeTrim prompt costs
Completion tokens spikingResponses run longer than they need toAdd a length instruction to the prompt, validate the candidate in Testing, then save it as a variant
Call count spikingRetry loop, crawler, or one abusive userSpot the pattern in the request list’s session and user columns, then fix the code or policy
Expensive modelModel overkill on trivial workSwitch to a cheaper model

Whether it is a prompt or a model, the fix follows the same order: confirm quality holds on the same inputs in Testing, save it as a variant, and raise its ratio in Routing.

Whether the savings actually landed shows on the cost trend of Overview’s 30-day range. Opening that screen once a month is enough to catch the next spike before the invoice does.