kadari client, rendered from an invented e-commerce workload — not from anyone's traffic. The arithmetic and the published rates are real; the calls are not. Run it on your own log to get your own numbers.Across 486,020 of 511,717 calls (95% of your log) that we hold a published price for. The other 25,697 are counted but not costed — their dollars are excluded from this figure rather than counted as zero.
| Model | Provider | Calls | In | Cached | Writes | Out | Spend |
|---|---|---|---|---|---|---|---|
| gpt-5.6-sol | openai | 441,031 | 467.5M | 894M | 0 | 8.4M | $3,037 |
| gpt-5.4-mini | openai | 44,989 | 35.1M | 0 | 0 | 247.6k | $27.41 |
The priciest 1% of your calls (4,860) account for 10% of the bill.
| Slice | Calls | Share of spend |
|---|---|---|
| Top 1% | 4,860 | 10% |
| Top 5% | 24,301 | 21% |
| Top 10% | 48,602 | 31% |
| Top 25% | 121,505 | 57% |
| Top 50% | 243,010 | 77% |
You spent $3,064 on calls that have a smaller rung in the same provider family. The identical tokens at that rung's published rate would be $128.79 — a difference of $2,935. Those calls are all of the spend we could price.
ESTIMATE from published list rates and your token counts -- not a saving, and not a claim that the smaller model would have been acceptable on your task. That is the question Kadari measures; this is only its size.
| Family | Cheapest rung | Calls | Paid | At that rung |
|---|---|---|---|---|
| gpt-5.6 family | gpt-5.6-luna | 441,031 | $3,037 | $121.47 |
| gpt-5.5 family | gpt-5.4-nano | 44,989 | $27.41 | $7.32 |
That difference is the ceiling on what routing could ever save you. What it does not tell you is how much of it you could take without your output getting worse — because nothing in this file has looked at whether a smaller model would have given an acceptable answer on your task.
That is the question Kadari measures. We re-run a small, cost-capped sample of your calls on the smaller rung and check the answers against the ones you already paid for — so the number you get back is one we can show the working for, not one we modelled.
Send us the log and we will do it: https://kadari.ai/submit.
Worth knowing before you do: the first report's proven savings figure is $0.00 by design. Until enough of your own calls have been checked, we have not earned the right to claim a number on your task, and we would rather say so than estimate one.
We hold no published price for the models below, so their calls are counted in the volume figures and excluded from every dollar figure. If you know the rate, pass your own table with --prices.
| Model | Calls | In | Out |
|---|---|---|---|
| ft:gpt-5.4-mini:acme-support:9RtQk2 | 25,697 | 20.6M | 141.1k |
Rendered locally by kadari/0.3.0 from your own log. No network, no JavaScript, nothing uploaded.
openai rates checked 2026-08-05 · anthropic rates checked 2026-08-05
Prices are a dated snapshot and decay; re-check before relying on a figure.