Savings estimate · prices checked 2026-11-03
GPT-5 mini vs Levanto Sage 2 for classification calls
List prices
| Model | $ per M input | $ per M output | Source |
|---|---|---|---|
| GPT-5 mini | $0.25 | $2 | OpenAI pricing |
| GPT-6 Luna | $0.1 | $0.5 | OpenAI pricing |
| Levanto Sage 2 | $0.05 | $1 | Levanto list price |
Checked 2026-11-03. Same table the public prices endpoint serves. Prices change; the audit always uses the current table.
Cost per call: our assumptions
- 500 input tokens per call (a ~100-token message plus a 400-token prompt, the audit’s defaults).
- 5 visible output tokens (“yes”/“no” plus formatting).
- 120 reasoning tokens at GPT-5 mini’s API default effort (
medium), billed as output. This is our estimate, not a provider figure. - Levanto Sage 2 charged the same 500 input tokens (conservative: it needs no instruction prompt) and 2 output tokens, since it returns a decision and a probability.
That gives an estimated $0.375 per 1,000 calls on GPT-5 mini, against $0.027 on Levanto Sage 2.
Effort level changes the bill
Reasoning tokens are billed at the output price, so the effort setting can change GPT-5 mini’s cost a lot. Estimated cost per 1,000 calls by effort level:
| Effort | Est. reasoning tokens | Per 1,000 calls |
|---|---|---|
| minimal | 10 | $0.155 |
| low | 40 | $0.215 |
| medium (API default) | 120 | $0.375 |
| high | 320 | $0.775 |
If a lower effort still gives the answers you need, that is a saving with no new vendor. Test it the same way.
Monthly estimate by volume
“Share kept” is a scenario, not a prediction: the share of calls where Levanto Sage 2 is confident enough to keep. The rest go to GPT-5 mini too, so the hybrid pays for both on those calls.
| Calls / month | All on GPT-5 mini | Hybrid, 50% kept | 70% kept | 90% kept |
|---|---|---|---|---|
| 10,000 | $3.75 | $2.15 | $1.40 | $0.645 |
| 100,000 | $37.50 | $21.45 | $13.95 | $6.45 |
| 1,000,000 | $375 | $215 | $140 | $64.50 |
Estimates at the assumptions above. A full switch to Levanto Sage 2 would cost $2.70 at 100,000 calls, but only makes sense if it agrees with GPT-5 mini on nearly everything.
When the switch is worth it
The audit only recommends a hybrid when it is at least 30% cheaper than staying put. With these prices that needs Levanto Sage 2 to keep at least 37% of calls at your agreement target. Below that, the saving rarely pays for the change.
- Good fit: the same yes/no or labelling prompt runs thousands of times (lead qualification, ticket triage, spam or intent checks, field extraction).
- Keep GPT-5 mini: free-text answers, tool use, multi-step reasoning, or inputs that differ a lot call to call.
Check it on your own calls
Paste 10 to 25 inputs with the answers GPT-5 mini already gave into the free cutover audit. It measures agreement, picks the confidence threshold that holds your target and projects the cost with your token counts. How it works · Choosing a threshold · Other model pairs · Reasoning effort cost