Guides / llm cost in dynatrace
How to see LLM cost in Dynatrace
Dynatrace AI Observability assesses token usage and cost across models and says it can predict cost increases so you can act before they land. Culpa, a local-first LLM cost, margin, and forecast ledger, projects next month from your own history as a range and sets each customer's spend against the revenue they pay, which is a different question from spotting a rise.
Why this happens
Dynatrace goes further on cost than most of this tier and the exact wording matters. Its AI Observability page says it assesses token usage, cost, stability, latency, invocation errors and resource utilization of model outputs, and that you can monitor operational metrics like token cost alongside request duration. It then says you can use intelligent detection to identify changes in user behaviour, predict cost increases, and proactively make changes to manage costs. Its documentation lists cost as token usage, service fees or overall resource consumption, and suggests error budgets for cost control. Read that carefully and it's detection with foresight: something is trending the wrong way and here it comes early. That's genuinely useful and it isn't a spend projection. A forecast answers what next month costs, with a range wide enough to plan against, derived from your own history. Predicting an increase answers whether something is about to get worse. Both are worth having and only one of them goes in a budget.
What this usually looks like
- You get told costs are rising and still can't say what the month will total.
- An alert fired on a cost trend and the plan it threatened was never a range.
- Token cost is visible per model and not per customer.
- Cost control is expressed as an error budget rather than as a margin.
- Nobody can say which customer's behaviour changed when the detector fired.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Mistakes that cost the most
| Mistake | Why it hurts | Do instead |
|---|---|---|
| Reading a cost-increase prediction as a spend forecast. | One tells you something is trending up, the other tells you where you land. Only the second is plannable. | Keep the detector for early warning and project the month separately, with a range. |
| Assuming a general platform can't see token cost. | Dynatrace publishes token cost assessment across models, so the gap isn't visibility. | Use what it shows and be specific about the question it doesn't answer. |
| Setting an error budget for cost without a revenue figure. | A budget says what you'll tolerate. It says nothing about whether the spend was worth it. | Set spend against the revenue it produced, then decide what the budget should be. |
| Acting on a detected rise without attributing it. | A rise caused by a new customer and one caused by a prompt regression need opposite responses. | Attribute the change to a customer, feature or prompt version before responding to it. |
Run this check tonight
- Find where token cost appears per model, then try to find it per customer.
- Ask what next month costs, as a range, and see which system can answer.
- Check what your last cost alert was actually caused by, and how long that took to establish.
- Write down what your cost error budget is protecting, in revenue terms.
A detected rise and a projected month answer different questions
A modelled workload growing steadily, priced on Claude Haiku 4.5 at real rates of $1.00 and $5.00 per million from the price book effective 2026-07-02. A detector fires when the rate of spend changes materially. A projection instead asks where the month lands. Both are run on the same numbers so the difference is the question rather than the data. Volumes are modelled.
The detector was right and it never produced a number anybody could budget against. The projection gives a range with a floor and a ceiling, which is the form a finance conversation needs.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| $312.00 to $426.24 | modelled monthly range for the same workload, flat against continued growth | estimated | Both bounds run from the same two observed weeks, $60.00 then $84.00, priced on Claude Haiku 4.5 at real rates of $1.00 and $5.00 per million from the price book effective 2026-07-02. If growth stops after week 2 the month is $60.00 + $84.00 + $84.00 + $84.00 = $312.00. If 40% weekly growth continues the month reaches $60.00 + $84.00 + $117.60 + $164.64 = $426.24. Every volume is modelled and the figure is published as a range because it's an estimate. |
What a generic answer can’t know
Dynatrace watches a system and reports when its behaviour changes, which is what an observability platform is for and it does it across token cost as well as latency. The two things it has no way to produce are the ones that need data it never receives. A projection of next month needs your own priced history held long enough to project from. A margin needs the revenue each customer pays you, which lives in a billing system rather than in telemetry. Culpa holds both, prices every call from a versioned price book in exact decimal, keeps the ledger on your own infrastructure so history length is a disk decision, and publishes the forecast as a range because a single number hides how much your own months disagree.
Questions founders ask next
Does Dynatrace track LLM cost?
Yes. Its AI Observability page says it assesses token usage, cost, stability, latency, invocation errors and resource utilization of model outputs, and its docs describe cost as token usage, service fees or overall resource consumption. Visibility isn't the gap here.
Does Dynatrace forecast LLM spend?
It says you can predict cost increases and act proactively, which is detection with foresight rather than a spend projection. A forecast answers where the month lands, with a range, from your own history. Both are useful and only one goes into a budget.
What's the difference between a cost alert and a forecast?
An alert tells you something changed, after it changed, no matter how early. A forecast tells you where you'll land if nothing changes, before it happens, with bounds. Teams that have the first often assume they have the second, which is how a month arrives 40% over plan with every alert green.
Can Dynatrace tell me if a customer is profitable?
No, and not because of any deficiency in the product. Profitability needs the revenue that customer pays you, which lives in your billing system and never reaches a telemetry platform. That join is the thing Culpa exists to make.
On your infrastructure
Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.
How Culpa works
Find the culprit. Not just the total.
Your dashboard shows what you spent. It stops short of who spent it. Culpa shows the conversation, the user and the feature behind it.
Your prompts stay local.
Culpa runs on your own infrastructure. What you send to a model reaches us at no point.
Every dollar has a name.
Follow any charge to the conversation, the user, the feature and the customer behind it.
See the bill before it lands.
Cost your next feature before you ship it. You get the likely bill and the worst case, at best, median, p90 and p99.
Three steps to your first answer.
Change one base URL.
Or drop in the Python or TypeScript library.
Find your most expensive conversation.
In the first session, not the first week.
Cost your next feature before you ship it.
Why the bill went up
Example dashboardCalls traced
418,209
across 3 projects
Spend this week
$378.41
+ $182 vs last week
Failed calls
312
74% retried, and you paid for all of them
+ $182 this week traced to one culprit
Spend over 14 days
Most expensive users
Next week forecast
Graded against reality. Accuracy shown as results land.
Keep reading
Sources: Dynatrace AI Observability, Dynatrace AI observability docs. Last reviewed 2026-08-03, rates effective 2026-07-02. Plain text version.