Understanding GPT-3 pricing is not a straightforward task since it has multiple base and trained models when it comes to API integration. In this article, I have simplified the prices of OpenAI’s GPT-3 models.
OpenAI has offered pricing for four base models named, Ada, Babbage, Curie, and Davinci. These models have different capabilities; hence the different price points. Out of these four base models, Ada is the fastest model and Davinci is the most powerful model. The four base models can be customized as fine-tuned models or embedding models according to user requirements.
Before going further into GPT-3 enterprise pricing, you must know what are tokens.
Tokens in GPT-3 API pricing are the representation of words. 1 token represents 4 characters/ 0.75 words approximately for English text. 1k tokens are equivalent to 750 words.
Choosing the Model – API
Although GPT-3 has simple and flexible pricing, the real struggle comes in while choosing the models. Let’s explore the models one by one.
Before that, those who want to use the GPT-3 API, need to submit the pre-launch review of their application. The purpose of this is to ensure that OpenAI’s policies and safety requirements are followed. For more information on who is required to apply this and what is required for approval, and how long the process takes, check OpenAI documentation.
Base Models
OpenAI has implemented the concept of “pay as you go” to keep things simple and flexible. The company also offers $18 in free credits that must be used within the first 3 months.
The models in the image are just the base models and there are custom models also available. Base models are pre-trained with billions of parameters.
Ada vs Babbage vs Curie vs Davinci
| Models (ordered by latest to old) | Features | Request Limit | Training Date |
| text-davinci-002 | Latest and most capable model. It can outperform other models with less context. Also, supports inserting text within the text. | 4K tokens | Up to June 2021 |
| text-curie-001 | Capable than Ada and Babbage models. Faster and lower cost than Davinci. | 2,048 tokens | Up to October 2019 |
| text-babbage-001 | Capable of uncomplicated tasks, very fast, and lower cost. | 2,048 tokens | Up to October 2019 |
| text-ada-001 | Good at very simple tasks, the fastest model with the lowest cost | 2,048 tokens | Up to October 2019 |
While Davinci is the superior model, Curie can also perform many of the same tasks as Davinci. You can also choose any model and improve its performance of it by fine-tuning them on a specific task.
Custom Models
Fine-tuned Models and Pricing
If you want to fine-tune a base model according to your custom needs, you can do so with separate prices.
Once you are done with fine-tuning the model, thereafter you will be paying only for the usage, not for both (training and usage).
Pros of Fine-tuned Models
- High-quality results than the base model
- Train on more examples
- Eliminates the need of giving examples in prompts. Hence, shorter prompts lead to token saving
- Low latency
Embedding Models and Pricing
In machine learning, embedding refers to using vectors to represent real-world objects and relationships. It is a dense representation of text that can be easily utilized by machine learning models and algorithms.
You are in need of an embedding model if you have to deal with massive amounts of data to train.
Pros of Embed Models
- Keeps data very simple for training and prediction
- Very low latency
Conclusion
OpenAI has a diverse range of pricing for GPT-3 API models. As mentioned in this article, these models also can be customized according to the user’s needs.
Also, these APIs have in-built free content filtering, end-user monitoring to prevent misuse, and endpoints to scope API usage.