Question about training cost and user model

[StoKou]

Thank you so much for your excellent work. I noticed that the user in the TAU2 benchmark uses GPT-4-1, but you used GPT-5 for both training and evaluation. I'd like to ask if you've conducted any additional experiments to explore the impact of different users? It seems to me that this impact is significant for more difficult tasks. Could you explain the considerations behind choosing GPT-5? And could you provide a rough estimate of the training cost? This is very important for us, given our limited resources. Thank you again for your outstanding open-source work!

[hongjin-su]

Hi, we used GPT-5 because we believed that stronger model would better simulate users, and thus porvide more accurate results. We do not carefully compare with GPT-4.1 due to the time limit. The training cost would be high, which costs 32 H100 more than a week to complete.

[mx2907]

Hi, we used GPT-5 because we believed that stronger model would better simulate users, and thus porvide more accurate results. We do not carefully compare with GPT-4.1 due to the time limit. The training cost would be high, which costs 32 H100 more than a week to complete.

What about the overall cost of api calls during the training i.e how many dollars needed to replicate the same training?