Guides / deep trace

What is a deep trace?

A deep trace opens one conversation all the way down to the prompt, so you can read what actually happened rather than what a chart summarised. Culpa, a local-first LLM cost, margin, and forecast ledger, meters these separately from ordinary tracking because they're the expensive question, and an entry plan carries a few hundred a month.

Why this happens

Two different jobs sit inside cost forensics, and they price differently for good reason. Recording every call is cheap and continuous, so it's metered in millions. Opening one conversation to the prompt is a deliberate investigation, so it's metered in hundreds. That scarcity is the useful part, because it forces the question nobody asks of an unlimited tool: which conversation is worth reading. Aimed at the top of the cost distribution a few hundred traces cover most of the money. Aimed at whatever alerted most recently they cover noise.

What this usually looks like

  • Traces get opened on whatever alerted, rather than on whatever cost the most.
  • Nobody has ranked conversations by spend before deciding which to open.
  • The allowance runs out mid-month and the expensive conversations were never read.
  • A single investigation opens dozens of traces to answer one question.

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

MistakeWhy it hurtsDo instead
Spending traces on alerts instead of on cost rank.An alert fires on a threshold, not on size, so the loudest call is rarely the dearest one.Sort by spend first and open traces from the top down.
Opening a trace before checking the aggregate view.Most questions resolve at the feature or model level for no forensic cost at all.Answer what you can from the cheap view, then spend a trace on what's left.
Treating the allowance as a monthly target to use up.Unused capacity isn't waste, and a month with few investigations is usually a good month.Track how much of your spend your traces covered, not how many you ran.

Run this check tonight

  1. Rank last month's conversations by cost and look at the top twenty.
  2. Work out what share of total spend those twenty carry. It's usually large.
  3. Count how many traces you opened, and how many were in that top group.
  4. Divide traces spent on the top decile by traces spent overall. That ratio is your aim.
  5. Check whether your allowance ran out before the expensive conversations were read.

Where a few hundred traces should point

Illustrative example

A product holding 180,000 conversations in a month against an entry allowance of 250 deep traces. The cost distribution below is modelled on the long tail these workloads usually follow.

250 traces against 180,000 conversations is 0.14% of them
Spent at random, 250 traces read 0.14% of the spend, which is nothing
Ranked by cost, the top 0.14% of conversations often carry a double-digit share of the bill
So the same 250 traces move from covering a rounding error to covering the expensive tail
The allowance never changed. Only the aim did

A scarce forensic budget is a targeting problem, not a capacity problem. Two hundred and fifty traces are plenty if they land on the top of the distribution and useless if they land anywhere else.

Every number, with its confidence and source

FigureWhat it meansConfidenceSource
250 to 1,200deep traces carried per month across the three published tiersestimatedCulpa's published plan table, read 2026-08-02, and the same figures in the shipped plan configuration. A range because it spans the entry and top tiers.
0.14%modelled share of a month's conversations an entry allowance can opencalculated250 deep traces divided by 180,000 modelled conversations, from the teardown arithmetic.

What a generic answer can’t know

Which conversation deserves reading is a question about your own cost distribution, and no provider holds that ranking. It needs every call priced and grouped by conversation first. Culpa does that on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.

Questions founders ask next

What does a deep trace show?

One conversation opened down to the prompt, so you can read the actual exchange that spent the money rather than an aggregate of it. That's the difference between knowing a feature is expensive and knowing why.

How many deep traces do I need a month?

Fewer than most teams expect, if they're aimed well. The published tiers carry 250, 500 and 1,200 a month, and a few hundred aimed at the top of the cost distribution covers most of the spend worth investigating.

Why meter traces separately from tracked calls?

Because they're different jobs at different scales. Recording every call is continuous and cheap, so it's counted in millions. Opening one conversation to the prompt is a deliberate investigation, so it's counted in hundreds.

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.

1

Change one base URL.

Or drop in the Python or TypeScript library.

2

Find your most expensive conversation.

In the first session, not the first week.

3

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

$0$20$40$60$8024262830020406
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.

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: Anthropic pricing. Last reviewed 2026-08-02. Plain text version.