What Bucket spent
Every judgement starts at the small model. The large one is reached only when the small one contests what a person said about their own money, which is an unusual reason to escalate: not that the task looked hard, but that a human disagreed. This page is the ledger behind that, read straight out of the table the app writes as it runs.
The split
- Purchases judged: 13
- Of those, escalated to the large model: 11 (85%)
- Escalations where the large model sided with the person and the objection was dropped: 0
That last line is what the expensive call bought on those occasions. Not a better answer shown to someone, but an argument withdrawn before they ever saw it. The individual cases are in the exchange log, which needs a key.
Per tier
| Tier | Calls | Tokens | Cost | Mean latency |
|---|---|---|---|---|
| Screening (Nemotron 3 Nano 30B) | 13 | 12,343 | $0.0012 | 4116ms |
| Adjudication (Nemotron 3 Ultra 550B) | 11 | 9,460 | $0.0189 | 1841ms |
Degraded calls are excluded: a fallback is not the model answering, and counting one would credit a tier that never ran. Cost is this app's own estimate rather than Nebius billing, because Token Factory does not return pricing with a response.
What routing was worth
- Screening actually cost: $0.0012
- The same screening at the large model's observed rate: $0.0247
- Difference: $0.0234 (19.9x)
- Total actually spent, every tier included: $0.0202
How that is calculated, so it can be argued with. The tokens the screening pass actually used are repriced at the cost per token the large model actually charged, both taken from the table above. It is deliberately not an average cost per call: the large model answers adjudications and reconsiderations, and a reconsideration carries the person's explanation and is a longer prompt, so a per-call average would inherit that length and overstate the saving.
What it assumes is that the large model, given the same prompt, would have written a similar-length answer. That is an assumption and not something this ledger can show. The escalation spending is left out of the comparison on purpose, because those calls happen either way.