Sample work · AI & Automation
AI Cost Analysis: Where an App Loses Money on Model Calls
From the call log of a sample app, it was determined where the costs for language models originate: duplicate requests, retries, overly large models, long prompts, missing output limits. Every proposal was checked for equal quality.
Sample work (demo), not a client job. Built as a trial task with invented example data; the check figures below come from the automatic test run and are not polished.
What was built
- Cost per function and model from the log, recalculated precisely
- Savings patterns identified: cache, retries, model selection, prompt caching, prompt shortening, output limit
- Every change tested against a quality check before counting as a saving
Download ki-kosten-auswertung.xlsx Download ki-kosten-auswertung.csv
Prices are sample values, not current list prices of any provider. There was no real model call in the sample; for the customer, calculations will be based on their real counters.
How it was checked
Automatic test run on 2026-10-31, taken from the log (test data invented, figures measured):
- Log 2000 calls/7 days, 171.96 USD (recalculated with decimals)
- Sample = Target: cache 310/310 (7.6% of costs), retry 30/30 (3.4% of costs), routing 718/718 (3.0% of costs), prompt_caching 631/631 (42.9% of costs), promptkuerzung 631/631 (42.9% of costs)
Limits
What this sample does not cover, so nobody expects more:
- Token counts are estimated (characters/4) where no actual provider counters are present in the log; at the customer's site, the usage fields of the API responses or the provider's tokenizer will be used
- Prices are sample values as of the reference date (preise.json) - no claim of current list prices; verify with the provider before any offer
- Prompt caching simplified: read price = 10% of input price, cache lifetime and write surcharge not modeled - actual savings depend on call frequency and provider
- Quality is measured here using deterministic model dummies with known quality; real quality requires the same test with real models and real customer tasks at the customer's site
- There is not a single model or paid API call in the sample
- After routing to a different model, the same output length is assumed; real models respond with varying lengths
- Cache only for exact identical inputs; semantic caching (similar questions) is not part of the sample
Something like this for your business?
Describe briefly what you need. You get a fixed price before I start; nothing happens without your yes.