ACLP

API cost leak profiler

Your AI bill is only telling half the story.

ACLP connects your LLM API spend to the features, users, and requests behind it, so you can understand what is driving costs and where to improve.

Explore the productGet started

The workspace below runs on a generated sample dataset. Nothing here contacts a provider or needs an account.

Northwind Labs · last 30 days

sample data

Total spend

$1,544.38+40.1%

Across 18,805 requests and 761.9M input tokens

Where it went

  • Document summarization$658.73
    43%
  • Code review assistant$634.94
    41%
  • Meeting notes$180.99
    12%
largest changeDocument summarization

Median input per request moved from 101,213 to 117,760 tokens while median output stayed flat.

Requests observed
791
Median input
117.8k tokens

The blind spot

You know what you spent. But do you know why?

A provider dashboard answers one question: how much. It cannot tell you which feature caused the increase, whether the work was necessary, or what to change. ACLP answers the second question.

Switch between the two views. The total never changes. The explanation does.

September · all providers

$1,544.38

  • OpenAI$1,378.85
  • Google$165.53

The total is 41% higher than the two weeks before. Nothing on this screen explains which change caused it.

The hidden cost

The expensive part is not always the obvious part.

Cost leaks rarely look like waste in a code review. They look like normal failure handling, normal re-renders, and normal conversation history. Four patterns account for most of what we see in sample datasets.

  • 01

    Retry loops

    A client timeout fires before the first response returns, so the same request is sent again and charged twice.

    retry gap: 6s · 31% of requests in this feature

  • 02

    Duplicate prompts

    A re-render or double click resubmits an identical prompt within the same minute, and only one result is kept.

    4+ identical prompts per user per day

  • 03

    Context growth

    Earlier turns are replayed in full on every call instead of being summarised once, so input grows with the conversation.

    median input ×1.9 · output unchanged

  • 04

    Model escalation

    A fallback route quietly moves routine traffic from an economy model to a flagship model with the same token counts.

    same tokens · 8× input price

The layer

Meet the layer that connects every request to its cost.

ACLP sits beside your application. It observes what the application already does, then answers the question your provider bill cannot.

  1. 01

    Application

    Your product code calls the model as it does today.

  2. 02

    SDK telemetry

    The SDK records usage metadata alongside each call.

  3. 03

    ACLP

    Events are normalised, priced, and attributed.

  4. 04

    Analytics

    Spend is broken down by feature, user, model, and session.

  5. 05

    Investigation

    A finding links the change to the requests behind it.

Observational by design
ACLP reads telemetry from your application. It does not intercept, proxy, or replace the provider calls your product already makes.
OpenAI and Google Gemini initially
The initial SDK covers OpenAI and Google Gemini. Additional providers are planned rather than shipped.
Prompt content stays out by default
The telemetry contract records token counts, model, latency, and request metadata. Full prompts and completions are excluded by default.

Interactive sample

Follow one cost spike from signal to next step.

Four steps, one dataset. Nothing is mocked up as a screenshot: the numbers below are calculated from the same sample ledger the product workspace uses.

sample data

Weekly spend for the customer platform

Eight weeks of sample telemetry with the previous period shown as a dashed line. One week breaks away from the pattern.

Peak week
14 Sep
Spend
$427.33
Change
+34.9%

The week of 2026-09-14 is +34.9% against the previous week. Select it to see what moved.

The workflow

From signal to explanation to action.

  1. 01

    Detect

    A spend change is separated from normal growth.

    Daily spend is compared with a trailing baseline for the same project and environment, so a real anomaly is not buried in expected growth.

    Daily spend 41% above the trailing 14-day average

  2. 02

    Understand

    The change is attributed to a feature, model, session, or request.

    Attribution narrows the signal until one or two drivers remain, then separates volume effects from behaviour effects.

    Document summarization · input tokens ×1.9, output unchanged

  3. 03

    Investigate

    Evidence points to specific requests and a concrete next step.

    Every finding keeps its supporting records, the measurement method, and an estimate that is labelled as an estimate.

    6 linked records · estimated excess labelled as modelled, not measured

Implementation

What this does, and what it does not.

Practical answers about telemetry, coverage, and estimates. Where something is not built yet, it is described that way.

More detail in the documentation.

Your application wraps its provider client with the ACLP SDK. Each call emits an event with usage metadata — model, token counts, latency, status, and the feature label you assign. Events are batched and sent to the ingest endpoint. This prototype does not send anything anywhere; the workspace is generated locally in the browser.

Get started

See what your AI spend is really doing.

Connect one service, attribute one week of traffic, and find out whether the increase you are looking at is growth or inefficiency.

Get startedReplay the sample investigation

Sample data — every number in this prototype is generated locally and is not real customer telemetry.