When usage climbs but the number of users and requests stayed flat, something has decoupled your cost from your traffic. This is one of the more revealing patterns in AI spend, because it points away from growth and toward a mechanism inside your own system. Flat traffic with rising cost is almost always one of the following.
Each interaction is consuming more tokens over time
The same number of users can generate a rising bill if each interaction costs more than it used to. In conversational and agentic features, this happens naturally as sessions accumulate history. If your average session length has been growing, the tokens per session grow with it, and cost rises even though the count of sessions is flat. Nothing broke. The unit cost drifted up under a stable volume.
Retries are inflating calls per request
Flat user traffic can hide rising machine traffic. If a rate limit, a provider slowdown, or a code change increased how often calls fail and retry, the number of API calls per user request climbs while the user-facing request count stays exactly the same. The users see no difference. The bill sees every retry.
An automation is running independent of user traffic
Not all AI cost is triggered by users. Scheduled jobs, background agents, and data processing pipelines make calls on their own cadence. If one of these scaled, or a new one launched, usage rises with zero change in user traffic, because the driver was never user traffic in the first place. This is why correlating spend only against user metrics can be misleading.
A prompt or context change raised the floor
A change that added tokens to every request, a longer system prompt, an extra retrieval step, a larger set of examples, raises cost across all traffic uniformly. Because it affects every call equally, it looks like a general increase rather than a spike, and it is easy to miss precisely because it does not concentrate anywhere.
How to find which one it is
Compute two ratios over time: calls per user request, and tokens per call. If calls per user request is rising, you have a retry or automation problem. If tokens per call is rising, you have a context or prompt problem. If user requests are flat but total calls are not, something non-user-driven is running. These three ratios separate the causes cleanly.
The reason this pattern is worth understanding is that it breaks the intuition that cost should track usage. In AI systems, cost tracks tokens and calls, and both can move independently of how many people are using the product. Watching the right ratios, not just the user count, is what keeps a flat-traffic cost climb from becoming a mystery.