If you have ChatGPT Plus, you have the choice between two different language models used for generating responses:
Both have advantages and disadvantages and are suitable for different use cases.
In this article, I'll briefly introduce GPT-3.5 and GPT-4 and give you recommendations on how to best use them:
- GPT-4 makes 40% fewer factual errors and understands complex instructions significantly better than GPT-3.5
- GPT-4o only costs 2.5x more than GPT-3.5 (instead of the previous 20x) and offers multimodal capabilities
- GPT-3.5 only makes sense for simple tasks, most users should switch to GPT-4o
1. What is GPT-3.5?
GPT-3.5 (also incorrectly called ChatGPT 3.5) is an LLM (language model) from OpenAI. It represents an improved version of GPT-3, which was released in June 2020.
More precisely, GPT-3.5 is not a single AI model, but consists of the language models code-davinci-002, text-davinci-002, and text-davinci-003.
GPT-3.5 is the default model of ChatGPT, available even to free users. You can try ChatGPT with GPT-3.5 for free here.
It's well suited for routine text work that is "everyday and not too demanding" and doesn't require ChatGPT plugins:
- Formatting (e.g., creating or modifying HTML tables or formatting a blog article)
- Changing between formal and informal tone
- Quick brainstorming
- Translations
- Changing capitalization
- Creating tables of contents
- Creating smaller code snippets
- etc.
The great advantage of GPT-3.5 lies in its speed. It generates responses noticeably faster than GPT-4.
However, GPT-4 is significantly more powerful:
2. What is GPT-4?
GPT-4 (also incorrectly called ChatGPT 4) is the flagship language model in ChatGPT. It has been available since mid-March 2023.
GPT-4 was trained with significantly more data than GPT-3.5. It can process more information and capture more complex relationships between words.
Since its release, I've tested it extensively and am impressed by the results. It performs significantly better than GPT-3.5 especially in the following tasks:
- Math and logic
- Data categorization
- Programming
- Text comprehension (e.g., understands idioms and complex instructions better)
Since November 2023, the "DALL·E", "Browsing", and "Advanced Data Analysis" (also called Code Interpreter) modes are also included in GPT-4. With DALL·E, you can generate images directly in ChatGPT.
With ChatGPT plugins, you can access third-party services and data and extend ChatGPT with new features. With Advanced Data Analysis, you can analyze documents, such as Excel files or text documents.
However, GPT-4 has three drawbacks:
First, as mentioned above, it can be slow. Second, errors occur more often with GPT-4, for example, OpenAI's servers may be overloaded and generation may abort. Third, the number of queries for GPT-4 is limited to 40 per day.
That's why I usually use GPT-3.5 for simple text generation tasks.
3. GPT-3.5 and GPT-4 Comparison
GPT-3.5 | GPT-4 | |
|---|---|---|
| Parameter Count | unknown | unknown |
| Training Data | unknown | unknown |
| Knowledge Cutoff Date | 09/2021 (in all variants) | 09/2021 – 12/2023 (in newest "gpt-4-0125-preview") |
| Release Date | March 15, 2022 | March 14, 2023 |
| Context Window (in tokens) | 4K to 16K | 8K to 128K |
| Features (in ChatGPT) | Text generation Code generation | Text generation Code generation Image generation (DALL·E 3) Data analysis Image recognition (GPT4-V) File upload GPTs |
