Guides / power user ai cost

Power user AI cost, and the workflow hiding inside it

Power user AI cost is what your heaviest accounts spend, and the useful question isn't how much but why. Most heavy accounts are heavy because of one workflow rather than general enthusiasm. Culpa, a local-first LLM cost, margin, and forecast ledger, attributes spend to the feature inside the account, so the answer names a workflow you can fix rather than a customer you'd rather not lose.

Why this happens

Once a heavy account gets identified, the reflex is commercial: raise the price, add a cap, have the conversation. That skips the question worth asking. Heavy accounts are rarely heavy across the board. They usually run one feature in a way nobody anticipated, most often a bulk operation with no change detection, so the same work gets redone on a schedule. The account looks like a pricing problem and turns out to be a product defect that every account is paying a smaller share of.

What this usually looks like

One account sits far above the rest and the only plan for it involves a pricing conversation.
Nobody has split that account's spend by feature, only by month.
The heavy account's usage has a schedule to it, which points at automation rather than at people.
Similar but smaller versions of the same pattern show up in other accounts once anyone looks.

Free, no card, no account

Run the free Cost Leak Scan

It shows your most expensive conversation before you install anything.

Run a free scan

Mistakes that cost the most

Mistake	Why it hurts	Do instead
Treating a heavy account as a pricing problem before decomposing it.	You raise a price on a customer whose spend was mostly waste your own product created.	Split their month by feature first. The commercial conversation gets much easier once you know the cause.
Assuming a power user is a scaled-up median user.	The feature mix usually differs, so scaling median behaviour understates the cost and misses the mechanism.	Compare the feature mix of your heaviest account against your median one, as shares rather than totals.
Fixing the account rather than the loop.	The same loop runs in every account at smaller scale, so the fix only reaches one of them.	Fix the workflow. The heavy account is where the defect became visible, not where it lives.

Run this check tonight

Take your single heaviest account and split last month's spend by feature. Read the shares.
Check whether the largest share is interactive or automated. Automated means a loop, and loops have settings.
For any bulk operation, ask what fraction of the inputs actually changed since the previous run.
Compare that account's feature shares against your median account's. Where they differ is your answer.
Check whether the same pattern runs in other accounts at a scale nobody has flagged yet.

Decomposing one heavy account's month

Illustrative example

The heaviest account on a product running GPT-5.4 at real rates of $0.0025 per 1k input and $0.015 output. The three workloads and their volumes below are modelled.

Bulk document re-run, 20,000 docs at 1.2k in and 150 out: (24,000 x $0.0025) + (3,000 x $0.015) = $60.00 + $45.00 = $105.00

Interactive chat, 1,800 turns at 9k in and 500 out: (16,200 x $0.0025) + (900 x $0.015) = $40.50 + $13.50 = $54.00

Background sync, 30,000 calls at 0.6k in and 80 out: (18,000 x $0.0025) + (2,400 x $0.015) = $45.00 + $36.00 = $81.00

Account total = $105.00 + $54.00 + $81.00 = $240.00

82% of the bulk re-run touched documents unchanged since the last run = $86.10, or 35.9% of the account

The heaviest account spent more than a third of its month re-reading files that hadn't changed. That's a missing change check in one job, and every other account is paying a smaller version of it.

Every number, with its confidence and source

Figure	What it means	Confidence	Source
$240.00	modelled monthly model spend for one heavy account across three workloads on GPT-5.4	calculated	$105.00 + $54.00 + $81.00 from the teardown arithmetic at real GPT-5.4 rates per 1k tokens from the price book, effective 2026-07-02. Workload volumes are modelled.
$86.10	modelled share of that account's month spent re-processing documents that hadn't changed	calculated	82% of the $105.00 bulk re-run, from the teardown arithmetic at real GPT-5.4 rates. The unchanged share is modelled.

What a generic answer can’t know

A provider can tell you an API key spent a lot. Which feature spent it, whether the work was new or repeated, and which account it served are facts your application holds and nobody else does. Culpa attributes them on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.

Questions founders ask next

Should I cap or reprice my heaviest account?

Decide after decomposing it. If most of the spend is a bulk job repeating unchanged work, the fix is in your product and the account was never the problem. Repricing first solves a symptom and leaves the mechanism running everywhere else.

How do I tell automated usage from human usage?

Look at the shape over time. People produce ragged, bursty patterns tied to working hours. Automation produces flat or regularly spaced volume, and a schedule visible in the data is the strongest signal that a loop rather than a person is spending.

Is a power user always unprofitable?

No, and often they're your best signal about where the product is going. It becomes a problem when nobody can name what they're doing, because then the exposure grows with adoption and the first measurement happens after the bill.

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.

Base URLhttp://localhost:4545/v1Your traffic keeps flowing if Culpa ever stops.

Why the bill went up

Example dashboard

Calls 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

user_384report_generatorconv_91fprompt_v1894,220 tokens3 retries$6.81

Most expensive users

user_384$38.42

user_119$21.07

user_562$14.90

user_204$8.30

user_871$5.10

Next week forecast

Best$180

Median$240

p90$310

p99$395

Graded against reality. Accuracy shown as results land.

Run the free Cost Leak Scan Start 14-day trial

Free, no card, no account

Run the free Cost Leak Scan

It shows your most expensive conversation before you install anything.

Run a free scan

Keep reading

Sources: OpenAI API pricing. Last reviewed 2026-08-01. Plain text version.