$23,117$433a month.
A support assistant’s LLM bill, cut with six levers applied one after another, and an eval that scores every answer before and after. The baseline is deliberately wasteful: for a typical $30K a month bill the estimator below shows 60-83% off, and a lean stack saves far less. Nothing here is invented, including the misses.
Quality: 76 of 81 checks pass after (78 of 81 before), 2 fewer than before, every miss shown.
Cost per request: $0.019264 → $0.000361
Real model calls on a fictional workload: 81 requests, scaled to 1.2M a month at Anthropic list prices. How we measured
Real model calls. Tokens measured on 81 requests, scaled to 1.2M a month, priced at public list prices. Rows are in cents so they add up to the total; the headline rounds to the dollar.
Where the money leaks
Six levers, applied one after another to the same workload. Each bar is the original bill: green is what is still paid after that lever, the bright grey is what that lever removed. Open one to see the real request behind it.
Cheaper only counts if the answers are right
The same eval ran on both pipelines: exact gold labels for the classifier, an LLM judge with a written rubric for answers, summaries and nightly digests. Quality after: 76 of 81 checks pass (78 of 81 before). Failures are shown, not averaged away.
Check your own bill
Enter what you spend and how your calls look. You get a range, and which levers apply. Nothing you type or drop here leaves your browser: the page makes no network calls with your data.
How we measured
One workload, run for real. These are the assumptions behind every number above.
- Scale. 81 real requests (37 support answers, 24 ticket classifications, 12 summaries, 8 nightly digests), scaled to 1.2M a month in the same mix. Tokens come from each call’s usage field and are priced at Anthropic list prices read on 2026-11-11.
- Caching. Assumes a 95% cache hit rate on a stable prefix of at least 1,024 tokens. Cache TTL and storage fees are not modelled.
- Baseline. Deliberately wasteful: one big reasoning model on every call, default reasoning effort, a long prompt and ten chunks every time. Hidden reasoning tokens were roughly 60% of its output tokens, estimated from the length of the visible text. A stack that already runs lean saves less.
- Quality. 76 of 81 checks pass after, 78 of 81 before. The five misses are listed above with the judge’s note, and the first optimized pass scored 59 of 81 before three eval-driven fixes.
- Scoring. Exact match where a gold label exists, otherwise one LLM judge with a written rubric. The judge is a measurement, not ground truth, and all 81 requests were both tuning set and eval, so 76 of 81 is an in-sample score.
Get this on your own bill
Your bill, measured the way this page measures Corvid Cloud: an eval from your own traffic, a ranked fix list with dollars per fix, and the code for the top fixes.
Check your bill free in your browser: the LLM Bill Checker
$1,900fixed fee · about one week
Cut Your AI Bill
- An eval built from your own traffic, so every change is scored on what your users actually send
- A ranked fix list with dollars per fix, measured on your workload
- The code or PR for the top fixes, written against your codebase
- A before and after report like the sample report built from this page’s run (PDF)
If verified 30-day savings are below the $1,900 fee, we refund the difference. Verified saving = (old cost per request − new cost per request) × requests served.
Start here, or go further
- Free 48-hour AI check$0Send your AI bill export. You get a one-page report on your top 3 problems. No call needed.
- Switch Models Safely$1,500Model being retired? One model and one use case moved over. Quality on the agreed test set doesn’t drop, or we keep working free.
- AI on call$1,500/moWeekly tests on real traffic, alerts when a model update breaks something, a monthly cost and quality report. Priority $2,500/mo. Cancel any month.
Or write to khoatran.mma26@gmail.com. We reply within 24 hours, Hanoi time (UTC+7). All offers