Updates to Fine Tuning API

Just got an email about new pricing/plan updates to the fine tuning API.

Think all of this is a great idea and very developer friendly.

4 Likes

Yes this is important subtext here, tomorrow I’ll do some math to see how much my previous fine tuning data attempts would cost me under the new plan.

https://beta.openai.com/docs/guides/fine-tuning/pricing

Currently there is no fee for fine-tuning the model itself, however we will be introducing a fee for this starting December 8th 2021 at 00:00 UTC. After this time, trained tokens will be billed at 50% the base model inference rate per 1k tokens (trained tokens = tokens in file * epochs).

I understand this as:
1st epoch of model: 50% of current rate
2nd epoch of model: 100% of current rate
3rd epoch of model: 150% of current rate
4th epoch of model: 200% of current rate

If true than if it’s only fined-tuned for the first epoch than we would pay only half the cost of base model.

1 Like

We’ll clarify this in the documentation.

If you were to fine-tune a Davinci model on the dataset of size 100k tokens for 4 epochs, you would pay 100 * ($0.06 / 2) * 4 = $12

3 Likes

How to get access for Davinci fine-tuning?

1 Like

Follow the instructions in the documentation.