Reference25 usable modelsUpdated 15 October 2026
Hosting and terms
Where each usable model ran in the tests, and on what terms.
Every model here was run on a route with zero data retention, except Qwen 3.8 Flash, which is only on Alibaba's servers. EU hosting will be added where it can be confirmed.
| Model | Weights | Developer based in | Route in the tests | Zero data retention | Watch for | List price per million tokens, in / out |
|---|
List prices are OpenRouter prices on the route each model ran on, read on 15 October 2026. A model is charged for the tokens sent in and the tokens it writes back, so two models with similar list prices can cost very different amounts per statement. The test costs used September and October 2026 prices, so a few of them reflect older list prices. What each model actually cost per statement is in the full cost calculator in the S1 paper.
The 25 usable models from S1. The 9 that weren't usable are left out. How each model scores is on each study page, and an estimated full cost, including assumed checking time and mistakes, is worked through in the S1 article.
Changes
- 15 Oct 2026OpenRouter list prices per million tokens for the 25 usable models, on the route each ran on.
- 14 Oct 2026First version: cost per task, hosting route and zero data retention for the 25 usable models.