When OpenAI introduced GPT-3, its third-generation autoregressive language model, which uses deep learning to generate human-like text, it completely changed the AI game.
Through a set of APIs, developers got access to one of the world’s most sophisticated AI text generation engines at a reasonable price. The availability of GPT-3 gave birth to an entire category of AI writing assistants, like Jasper and Rytr, and enabled language learning apps, such as Duolingo, to transform their operations.
To calculate (or at least estimate) how much it will cost to use OpenAI GPT-3 for your application, you need to determine two things:
- Whether you will use base models or fine-tuned models
- Which model (Ada, Babbage, Curie, or Davinci) you will use
In this guide, I’ll give you the rundown and point you to the pages on OpenAI’s website where you can learn more, so read on!
OpenAI Base Models
OpenAI has four GPT-3 base models: Ada, Babbage, Curie, and Davinci.
Ada is the fastest, least capable, and cheapest. Davinci is the slowest, most capable, and most expensive. Babbage and Curie are respectfully in the middle in terms of speed and power.
| Model | Traits | Usage Price |
|---|---|---|
| Ada (Base) | Fastest, least capable | $0.0008/ 1K tokens |
| Babbage (Base) | Second-fastest | $0.0012/ 1K tokens |
| Curie (Base) | Second-most capable | $0.006/ 1K tokens |
| Davinci (Base) | Most capable, slowest | $0.06/ 1K tokens |
So, what’s the difference?
OpenAI Ada is the fastest and least capable base model. At $0.0008/ 1K tokens, it’s also the least expensive. Use it for basic tasks like parsing texts, extracting keywords, and correcting addresses.
OpenAI Babbage is the second-fastest and the second least expensive base model at $0.0012/ 1K tokens. Use it for semantic classification of entities that requires more brainpower.
OpenAI Curie is the second most capable and the second most expensive base model at $0.006/ 1K tokens. Use it for summarizing and translating text and basic text generation.
OpenAI Davinci is the most capable and slowest base model. It can perform any task that the other models can, usually with less training data and/or instructions. At $0.06/ 1K tokens, it’s also the most expensive base model. Use it to interpret and deliver complex intent and for advanced text generation.
And which one should you use for your AI-powered application’s needs?
Generally, OpenAI recommends using Davinci when experimenting because it gives the best results.
Once you’ve nailed the inputs and outputs, you can try the other models and see which is the fastest and most economical without compromising on the quality of the outputs. (Remember that you can improve the performance of any GPT-3 model by fine-tuning it with training data.)
OpenAI Fine-Tuned Models
OpenAI lets you fine-tune each GPT-3 base model with your training data. The cost of training a model is 50% of that base model’s usage rate.
| Model | Traits | Training Rate | Usage Rate |
|---|---|---|---|
| Ada (Fine-tuned) | Fastest, least capable | $0.0004/ 1K tokens | $0.0016/1K tokens |
| Babbage (Fine-tuned) | Second-fastest | $0.0006/ 1K tokens | $0.0024/1K tokens |
| Curie (Fine-tuned) | Second-most capable | $0.003/ 1K tokens | $0.012/1K tokens |
| Davinci (Fine-tuned) | Most capable, slowest | $0.03/ 1K tokens | $0.12/1K tokens |
Once you train a model, OpenAI’s website says, you will be billed at the usage rate for the fine-tuned version of that model.
Note that, although the training rate for a fine-tuned model is 50% of the base model’s usage rate, the usage rate for the fine-tuned model is higher than that of the base model.
If you want to use a fine-tuned GPT-3 model for your application—meaning one that you’ve trained with your own datasets—you need to calculate your API usage costs using the fine-tuned model’s rates significantly change the economics of your application.
How You’re Billed
When you create your account, you get $18 in free credit, which you will need to use up within your first 3 months.
This is enough for 22,500 credits running GPT-3’s Ada base model or 300 credits running the Davinci base model. (According to OpenAI, a thousand API credits is roughly 750 words.)
Once you’ve used up all of your free credits, you will be billed on a pay-as-you-go basis. Of course, if you work for a fast-growing startup or large corporation, you probably want to get your Procurement team involved to see if they can negotiate a deal with OpenAI’s sales team.