Fine-tuning Beta Release

Hi all! As you may have noticed, we have enabled fine-tuning for all users who have API access.

The feature, which is in beta, allows you to get more out of the models available through the API.

Fine-tuning will provide higher quality results than prompt design, enable users to train on more examples than can fit in a prompt, and result in token savings due to shorter prompts.

Currently, users can fine-tune Ada, Babbage, and Curie in a self-serve manner. Details can be found in the Fine-tuning guide.

We’re working to support fine-tuning Davinci and will provide access to users based on responses to this form.

Pricing follows the familiar pay-as-you-go structure as our base models. Pay-as-you-go rates can be found in the fine-tuning documentation

If you have any questions, issues, or general feedback, or if you would like to use fine-tuning for academic or research purposes, please contact


happy dance happy dance happy dance


This is super cool. Thanks so much.

Might be dumb question but regarding the pricing;

  1. It is free to finetune right and then we get the model ID and can use that.
  2. Once we call the Completions API on that model, we are charged the pricing shown but due to the smaller inputs because of finetuning, we should use less tokens overall?

Excited to play with this.

Awesome. Fine-tuning is a critical function that I outline in my AGI book - which will hopefully be released this coming weekend. Recording data is one way that AGI can learning, but fine-tuning itself as it accumulates data is another way. This ability will be a critical step!

1 Like

This is great. Thank you for your continued amazing work.
Can a model be fine-tuned over and over?
I mean, let’s say I need to fine-tune the language style based on specific literature but need to have the actual completion based on another set of data but in the same style.
Another example is training the model to improve its abilities in foreign languages. Still, we don’t want the model to base the completion on the materials used for learning the language but based on limited professional materials.
Can that be achieved?

1 Like

I wrote an article about it here Beyond Few-Shot Learning: Fine-tuning with GPT-3 | by Coleman Hindes | Jul, 2021 | Medium

  1. It is free to finetune right and then we get the model ID and can use that.
  1. Once we call the Completions API on that model, we are charged the pricing shown but due to the smaller inputs because of finetuning, we should use less tokens overall?

Yes! We’re charging for inference, not the training process. With enough training data you will no longer need to use few shot examples in the prompt, reducing the number of credits used per request.


Will it ever be possible to make a fine-tuned model available to other OpenAI customers in exchange for either pay or for free?



Can a model be fine-tuned over and over?

You can train new models as you get more data, but we do not currently have the ability to continue to fine-tune a model that has already been fine-tuned once. That is to say, every time you fine-tune you will start with a base model.

1 Like

It’s definitely possible, but our main focus right now will be making the endpoint as effective and easy-to-use as possible. Would love to know more about what you’re envisioning though! Can you share more on what you’d like to see and why?

1 Like

Yeah, no rush.

We’ve got nearly 10,000 tagged generations of items, people, monsters, locations, and more at LitRPG Adventures. I’m fairly certain it would work well as a general RPG/fantasy/scifi/TTRPG type of model, and I think others might be interested in using it. Might be another way for me to generate revenue to bootstrap my project by leveraging all the work I’ve already put in.

Just something I’ve been thinking about for a while, so I thought I’d plant a seed now.

Thanks for the quick reply!

1 Like

Makes sense! How would you think about competitive advantage for your project if sharing the fine-tuned model?


Good point. I’d likely want control over who gets access to the fine-tuned RPG model so as to align with companies / people that are more cooperative than competitive? (ie a writer or TTRPG game developer who wants the model for building stuff to sell rather than building what I’m attempting. The revenue generated might be more than what I’m getting now with customers, but then again, I might be shooting myself in the foot. Haha.

Definitely something to think about!

I’m super excited to fine-tune and see if I can get Curie to the same level as Davinci. That would allow me to offer a lot more value to my subscribers/members by getting the costs down. Might be able to go truly FREEMIUM at that point too…

Thanks again.


I remain very interested in OpenAI’s position on possible use of third-party owned IP in the fine-tuning datasets, and the possible leaks that could follow

Could you share an example of what you’re thinking of?

1 Like

Thinking out loud here: What if an end user takes something they don’t have the right to redistribute - let’s say they have a library of epub books that they’ve stripped the DRM off of. These books are not in the public domain but the user wants to finetune a model on this particular genre to aid with their own writing project. What liability would OpenAI place on the user? What would the legality of DMCA be here? Would OpenAI be compelled to delete that finetuned model? Or would this be protected under something like Section 230?

(Please note: I have no interest in doing this kind of thing - I was merely inspired to ask by the other thread about using GPT-3 to produce fiction, so this example was top of mind)


I wasn’t crystal clear after reading the guide: Prompt + completion for each item in the training data set is still limited to 2048 tokens?

1 Like

We have production approval and our use case fits best first fine tune the model before applying the completions api…do we need to see again production approval for fine tuning process or this can be considered against the existing use case…NOTE: we are not changing any part of the feature or application its just overlaying the finetuning process on top of it. Please confirm.


During the early stages of this beta, we’d love everyone to reach out when ready for production. Both to check-in on any safety questions, but also because we want to get as much feedback as possible so that we can continue to improve the product. We’ll follow up directly.


Rather it seams that engine_id will go and a model parameter will come across all endpoints. Not a bad idea, if made available immediately.

That’s right! To make sure I understand, what do you mean by “if made available immediately”? The model you trained?

1 Like

By production, do you mean after training has been completed and tested, but before releasing to end users?