Your AI bill is leaking.

Send a sanitized usage export.
In 48 hours, Traceminder shows where the money is going, how much is wasted, and what to fix first.

Request export check See sample report

How it works

1 Send the usage file

Share cost and usage fields only. No prompts, API keys, secrets, or customer names.

2 Traceminder finds waste

It shows where AI spend is going and which parts cost too much.

3 Fix the first leak

The report says what to fix first, with evidence your team can check.

Delivery Cost leak report in 48 hours
Input Sanitized usage metadata only
Privacy Your file is deleted after delivery

Upload

Share usage metadata and keep private data out

Traceminder only needs sanitized cost metadata from OpenAI, LiteLLM, Langfuse, Helicone, or your own logs.

Useful fields
timestamp model input_tokens output_tokens cost endpoint status retry_count cache_hit user_hash tenant_hash
Private data stays out
prompts completions API keys secrets customer names raw requests raw responses

Review the sample report before the file check. Traceminder works from sanitized metadata fields and keeps prompts, completions, API keys, and customer data out of scope.

Download CSV template View sample report
top_leaks Top cost leaks

See which routes, models, tenants, retries, or endpoints are driving spend.

estimate_waste Estimated waste

Understand how much each issue may be costing per month.

fix_list Fix list

Get practical changes your team can make first.

shareable Shareable report

Receive a clean PDF or HTML report you can send to engineering or product.

Methodology

How Traceminder finds cost leaks

Retry waste

Repeated failed or retried calls that still cost money.

Prompt bloat

Large static prompts or long contexts increasing every request.

Wrong model routing

Expensive models used where cheaper models may be enough.

Tenant outliers

Users or tenants spending far more than normal.

Cache misses

Repeated similar calls that could be cached.

Where spend leaks

How leaks show up in your traces.

Sample report

See exactly what you get.

A report is a short list of leaks. Each leak shows the wasted money, the proof in the data, and the first fix to make.

Open sample report
Example finding
Retry waste: $312/month

Failed calls are being paid for again because retries keep running.

Proof

184 repeated billed attempts appear in the usage file.

First fix

Cap retries and add backoff before the next bill grows.

traceminder-sample-report.pdf Open in new tab Download PDF
Sample report · demo usage data

Cost leak report

Based on 426 logged calls over roughly 0.9 days, projected to 30 days.

projected_monthly_spend $25.49
estimated_monthly_waste $16.60 About 65% of spend, de-duplicated so nothing is counted twice.
  1. Cache misses $11.10 / mo

    Identical requests repeat while cache_hit stays false, so each repeat is paid for.

    Evidence: /answer has 66 calls sharing only 4 input sizes, cache-hit rate 0%.

    Fix: Add caching keyed on the normalized request, enable provider prompt caching, and set sensible TTLs.

    Confidence: medium
  2. Tenant outliers $9.10 / mo

    A small number of tenants drive a disproportionate share of spend.

    Evidence: One tenant drives 55% of spend ($0.41), while the median tenant is $0.10.

    Fix: Review limits and pricing for these tenants, apply the fixes above to their traffic first, and add per-tenant budget alerts.

    Confidence: high
  3. Prompt bloat $5.34 / mo

    A large, near-constant prompt prefix is sent on every call to some endpoints.

    Evidence: /answer has 66 calls averaging 1315 tokens in and 249 out over a static floor of 1314 tokens.

    Fix: Trim the system prompt, move static context to provider prompt caching, and retrieve only relevant chunks.

    Confidence: medium
  4. Retry waste $3.10 / mo

    Calls are billed multiple times because failed or rejected responses are retried.

    Evidence: 108 billed retry attempts and 32 rate-limited or error attempts, most of them on /extract.

    Fix: Cap retries, add exponential backoff, fix the validation that rejects good responses, and alert on retry spikes.

    Confidence: high
  5. Wrong model routing $0.74 / mo

    Trivial tasks run on an expensive model where a smaller model is typically sufficient.

    Evidence: /classify has 110 calls on gpt-4o producing only 6 output tokens each.

    Fix: Route classification and short tasks to a mini model behind a quality check, and add a model-selection policy.

    Confidence: medium-high

Per-finding figures are standalone and can overlap. The headline waste attributes each billed call to a single leak, so it stays honest. Analysis uses only sanitized usage metadata. Prompt and completion text stay out of the report.

Products

Choose how deep Traceminder should look.

Start here Cost Leak Report $149

Shows where money is leaking, how much is wasted, and what to fix first.

Delivered in 48 hours as a PDF or HTML report.

Request export check
Go deeper Deep Cost Report $399

Includes the Cost Leak Report plus a deeper engineering plan.

Route review, cheaper model options, and a savings plan.

Request export check

Still not sure?

Ask before sending anything.

Tell Traceminder what you use today, and get a clear answer on whether a report makes sense for your team.

Email Traceminder

FAQ

Questions before you upload

How does Traceminder review my OpenAI account?

You share a sanitized CSV or JSON export. Account access stays with your team.

Which data is included?

Usage metadata such as model, token counts, cost, endpoint, status, retry count, cache flag, and hashed tenant or user IDs.

How fast is delivery?

Most reports are delivered within 48 hours after payment and usable data.

What if my file is missing some fields?

The file check returns the fields to add or a safer export option. The invoice is sent after the data is ready for a useful report.

Is this a dashboard?

Traceminder is a focused cost leak report that works alongside Langfuse, Helicone, LiteLLM, provider exports, or your own usage logs.

Field notes

Cost leak patterns Traceminder checks first

Practical notes on real LLM cost leaks: retry waste, prompt bloat, model routing mistakes, cache misses, and tenant outliers. Each note shows the data pattern, how Traceminder detects it, and what to change to reduce spend.

Review queue

Every report starts with retry storms, prompt bloat, expensive low-volume endpoints, cache misses, model routing mistakes, and tenant outliers.

Traceminder turns a sanitized usage export into a ranked cost leak report for teams running OpenAI, Claude, Gemini, RAG, or agent workflows.

See where the AI budget went.

Share usage metadata only. Traceminder checks that your file has enough data before invoice, delivers the report within 48 hours, and deletes the file after delivery.
Request export check