The AI meter billed some requests at twice the real price.
Twenty debits its customers AI credits from a model price catalog. The catalog sync gave every model the same long-context threshold, 200K tokens. Providers start that rate elsewhere: 272K for 13 OpenAI models, 100K for Claude Haiku 5.5.
So requests of 200–272K on those OpenAI models were debited at about twice the provider's price, and requests over 100K on Haiku 5.5 at about a fifth of it. The cause was a pricing field the upstream catalog deprecated in August 2026.
This is not a saving for Twenty. Customers were over-charged on some models and under-charged on one; the fix reads each model's real threshold so the meter matches the provider's price.
| model | input | provider price | debited |
|---|---|---|---|
| gpt-6-luna | 240K | $0.0051 | $0.0099 |
| gpt-6.1-sol | 240K | $0.0796 | $0.1542 |
| claude-haiku-5-5 | 150K | $0.0168 | $0.0034 |
The same thing in your product
If you resell AI usage to your customers, your meter is part of their bill. We check it against each provider's current price list, including the thresholds where the price changes.