There has been a lot of buzz recently about ChatGPT and for a good reason. ChatGPT does a fantastic job working with texts and is incredibly easy to use. So it immediately became viral. Also, it raised interest in the topic of Generative AI (artificial intelligence that generates content like text, images, audio, or video).

At the same time, many new terms and names popped up, sometimes creating confusion. Let’s see who is who (or what is what) in this field. This is important to understand, as these tools are different and should be applied differently.


The name of the company builds all the things listed below. It was founded in 2015, and among the cofounders were Elon Musk, Peter Thiel (Paypal, Palantir), Reid Hoffman (LinkedIn), and others. Also, in 2019 Microsoft joined the club, adding one billion dollars as investments and becoming an important technology partner.

This article will only focus on the things built by OpenAI. We may check some competitive products in the following posts.


A common name for AI models trained to work with text. Models (also called LLM or Large Language Models) are a fusion of mathematics and language, capable of interpreting and creating text. These models were trained for specific tasks (we will briefly cover the topic of model training in the following posts).

GPT-3 consists of four main models: davinci, curie, babbage, and ada.

All of them are available via paid API (Application Programming Interface). They are intended to be used by developers that want to enhance their products with AI. Also, using API allows more control over the behavior of the models compared to the latest ChatGPT.

Also, these models are available in Playground, a web interface that allows experimenting with the models without API.

These models have different characteristics and purposes. For example, davinci is the most capable model, which is best for writing text. However, it requires more computing power and therefore is the most expensive. On the other hand, ada is very fast and cheap and works best for simple tasks like text classification or model fine-tuning.


It is another model from the GPT-3 family. It stands out from the four models mentioned above. It was specially trained for dialogs, remembering previous discussions in each chat, asking follow-up questions, and adapting to the context. Also, it is free and easy to use by anyone. On March 1, 2023, the model behind ChatGPT called gpt-3.5-turbo became available via API.

ChatGPT hit the market and took it by storm. In the diagram below, you may see how fast did they acquire users:

Based on data from Similarweb

The other side of this success was unexpectedly high traffic and load on the servers. Very soon, the users started to experience system outages. And OpenAI had to respond.

ChatGPT Plus

On February 1, 2023, OpenAI presented a paid version of ChatGPT called ChatGPT plus. It offered the same functionality with high availability. When users of the free version received the errors caused by the increased server load, paid users enjoyed it without limitations.

Also, OpenAI promised the paid users early access to the new features. However, before March 14, 2023, it was unclear what features could justify it. Stay tuned, at the end of this article, you will see it ;-)

GPT-3.5 and InstructGPT

GPT-3 was introduced in 2021. Then in January 2022, some models were upgraded to a version called InstructGPT. That version changed the approach to sending instructions to the machine to plain English.

Later, OpenAI trained new and improved models that belong to the GPT-3.5 series. As of now, these are davinci and ChatGPT. However, many people, including OpenAI, still use the common name GPT-3.

GPT-3 does not work with anything else except text. However, OpenAI has more to offer:


It is a set of models that allow working with programming code. It also works with text but is specific to programming, not general-purpose, as GPT-3 does. These models are available via API and Playground. Also, they are available for ChatGPT, so you may ask it to write you a piece of code or interpret your code.


It is a model that generates images based on the text description. If you have heard about so-called AI Art, this is a good example. This model is available via API and directly on the website.


It is an open-source model which is created for Voice-To-Text. So you can transcribe audio content in different languages with an impressively low level of errors. This model is free and open-source. You need to download and install it, which requires technical skills. Furthermore, it will require some computing power. Starting March 1, 2023, you may also use it via paid API.


On March 14, 2023, OpenAI presented the next-generation model called GPT-4. We still need to evaluate it and see if it is significantly improved compared to earlier models, as many features are not publicly available. Nevertheless, here is the list of promises by OpenAI:

1) It is capable of working not only with text but also with images. As discussed above, OpenAI has DALL-E, which can generate images. And GPT-4 can analyze the images and create text based on that. For example, it can describe pictures with words;

2) It is more accurate to work with facts. “Hallucinations” of GPT-3 that resulted in wrong answers became frustrating for many users, and GPT-4 is promised to be better, yet not ideal;

3) It has a more extensive “context window.” When GPT-3 models allow processing from 2000 to 4,000 tokens in one request, GPT-4 processes from 8,000 to 32,000 tokens (please, check this article if you don’t know what tokens are)

4) It is supposed to be easier to steer, so the users and developers can receive more predictable results.

Space for improvement:

1) Even though this model is substantially younger than GPT-3, it was still based on the data with a cut-off in September 2021. It means the core model does not know anything about what happened after that date. And in some areas, like AI, where things change monthly, it may be frustrating;

2) It is more expensive compared to the davinci. As of March 14, davinci (not fine-tuned) costs $0.02 for 1,000 tokens (both for prompt and completion). And in GPT-4, the pricing model changed. Now they have two plans, depending on the context window size, and charge separately for the prompt and the completion. So, the price varies from $0.03 to $0.12 for 1,000 tokens.

3) It is not clear if we can fine-tune the model. For example, OpenAI currently does not allow fine-tuning their recent gpt-3.5-turbo, and GPT-4 may also have this limitation.

As of the day of the launch, this model is publicly available for ChatGPT Plus users.

This blog was originally posted by us on Medium