I’m working on a project that wants to finetune an individual model for each user, so this could go up to 60 and potentially scale to much larger than that. I know some posts mentioned that you can finetune up to 48 models per day, but does that mean you can own and use 144 models by day 3 or is it always a 48 limit? Does it cost money to keep unused models?
Appreciate the help in advance!!
OpenAI has never given a limit to the fine-tuning model count, and it does not cost any money to maintain them. You’ll just have a “list models” endpoint full of countless IDs.
It is likely your “per user” idea will not be fruitful or productive, however. The AI is trained by example responses of how it should respond to operate in a style and behavior, requiring development of an extensive training set of example conversations; fine-tuning is not to impart knowledge of user data.
Think about what is common to all users if an available model with prompting alone cannot do the task and you need fine-tuning.
It also simply costs more: more upfront costs and higher costs for individual calls. You can triple the prompt input with every call and still be paying less. Investment that lasts only until the base model is retired.
Thank you for your response. Why would it cost more for every call? Wouldn’t the extra cost only come from the upfront deployment? Please help me understand! Thanks.
It costs more to USE a fine-tuning model. For example,
| Model | Input | Cached input | Output |
|---|---|---|---|
| gpt-4.1 | $2.00 | $0.50 | $8.00 |
| ft:gpt-4.1 | $3.00 | $0.70 | $12.00 |
| gpt-4.1-mini | $0.40 | $0.10 | $1.60 |
| ft:gpt-4.1-mini | $0.80 | $0.20 | $3.20 |
And the training cost for gpt-4.1: $25.00 per million tokens, per epoch (pass)