GPT-3.5 is an improved version of GPT-3, offering better performance and understanding of human language. ChatGPT members use the GPT-3.5 model as their default model
Here, we will take a closer look at what OpenAI’s GPT-3.5 is and explain the differences between it, GPT-3, and the latest model, GPT-4.
OpenAI’s GPT-3.5 Model
The GPT-3.5 model is not a single model but a collection of different models optimized for various domains. There are a total of 7 different GPT-3.5 models available. These models serve different purposes and have varying token limits and training data.
OpenAI recommends using the gpt-3.5-turbo model among the other GPT-3.5 models due to its advantages of lower cost and improved performance. For reference, visit the related page.
Below, you can find details about each model:
| Model | Description | Maximum Tokens | Training Data |
|---|---|---|---|
| gpt-3.5-turbo | Optimized for chat applications, offering similar performance to text-davinci-003 at 1/10th the cost. | 4,096 | Up to Sep 2021 |
| gpt-3.5-turbo-16k | Same capabilities as gpt-3.5-turbo but with four times the context, allowing for longer conversations and larger input. | 16,384 | Up to Sep 2021 |
| gpt-3.5-turbo-0613 | A snapshot from June 13, 2023, with function calling data. No updates; deprecated after three months from a new version. | 4,096 | Up to Sep 2021 |
| gpt-3.5-turbo-16k-0613 | A snapshot from June 13, 2023. No updates; deprecated after three months from a new version. | 16,384 | Up to Sep 2021 |
| text-davinci-003 (Legacy) | Capable of language tasks with better quality, longer output, and more consistent instruction-following. Supports inserting text. | 4,097 | Up to Jun 2021 |
| text-davinci-002 (Legacy) | Similar capabilities to text-davinci-003 but trained using supervised fine-tuning instead of reinforcement learning. | 4,097 | Up to Jun 2021 |
| code-davinci-002 (Legacy) | Optimized for code-completion tasks, useful for programming-related queries. | 8,001 | Up to Jun 2021 |
The History of GPT-3.5 Model
OpenAI had announced the GPT-3 model on June 11, 2020. They introduced new versions of its AI models, GPT-3 and Codex, to their API on March 15, 2022 via a blog post. These models named “text-davinci-002” and “code-davinci-002” had enhanced capabilities for text editing and insertion of the current model.
On November 28, 2022, OpenAI released another AI model called “text-davinci-003.” Shortly after, on November 30, 2022, OpenAI began categorizing these models as part of the “GPT-3.5” series.
November 30, 2022, is also the date when ChatGPT was announced.
Therefore we can say that the GPT-3.5 model was first released on March 15, 2022. Its predecessor is the GPT-3 language model, and the subsequent model is the GPT-4 model. However, the it has been receiving updates since its initial release date.
To learn more, take a look at our post where we examine all GPT models of OpenAI.
Number of Parameters used to train GPT-3.5
The GPT-3.5 model was trained using 175 billion parameters, which were obtained from various sources. The distribution of these sources for training the model is as follows:
| Dataset | Percentage |
|---|---|
| Web Crawling | 60% |
| WebText2 | 22% |
| Books | 15% |
| Wikipedia | 3% |
These sources were utilized to gather a diverse range of data for training the model.
Character Limit in GPT-3.5 model
The GPT-3.5 model has a character limit. It can’t understand and provide unlimited responses.
Each model within GPT-3.5 has a different character limit, which is determined based on tokens. Tokens are a type of character, and roughly 100 tokens are equivalent to 75 English characters. You can find approximate character limits for each model below.
| Model | Character Limit |
|---|---|
| gpt-3.5-turbo | 3,072 |
| gpt-3.5-turbo-16k | 12,288 |
| gpt-3.5-turbo-0613 | 3,072 |
| gpt-3.5-turbo-16k-0613 | 12,288 |
| text-davinci-003 (Legacy) | 3,073 |
| text-davinci-002 (Legacy) | 3,073 |
| code-davinci-002 (Legacy) | 6,001 |
How to access and use GPT-3.5 model?
To access the GPT-3.5 model, there are three primary methods available:
- ChatGPT: Through ChatGPT, you can access to the GPT-3.5 model. The default model of ChatGPT is GPT-3.5. Plus members have access to both the GPT-3.5 and GPT-4 models.
- OpenAI Playground: You can access the GPT-3.5 model through the OpenAI Playground. It provides a user-friendly interface and allows you to interact with the model and customize various parameters without requiring extensive programming knowledge.
- API Integration: The third method is to call the GPT-3.5 model’s API. You can implement this in any programming language of your choice. However, there is a cost associated with each API call based on the number of tokens used.
Also Read: How to call GPT API?
Below is a simplified example of how to call the GPT-3.5 model in Python:
import openai
openai.api_key = 'your-api-key'
def call_gpt3(prompt, tokens=100):
response = openai.Completion.create(
engine="gpt-3.5-turbo",
prompt=prompt,
max_tokens=tokens
)
return response.choices[0].text.strip()
response = call_gpt3("Translate the following English text to French: '{Hello, how are you?}'", tokens=60)
print(response)
Remember to replace 'your-api-key' with your actual API key.
Price of Using GPT-3.5 API
The usage fee for the GPT-3.5 model is charged per token used.
Below are the usage fees for the Turbo model:
| Context | Input Rate | Output Rate |
|---|---|---|
| 4K | $0.0015/1,000 tokens | $0.002/1,000 tokens |
| 16K | $0.003/1,000 tokens | $0.004/1,000 tokens |
These prices are approximately 20 times lower compared to the usage of the GPT-4 API. For more information, take a look at OpenAI API pricing page.
GPT-3.5 vs GPT-3: What is the difference?
GPT-3.5 and GPT-3 are both models of artificial intelligence developed by OpenAI, designed for advanced natural language processing tasks. However, there are key differences between the two that position GPT-3.5 as a more evolved version of GPT-3.
ChatGPT is designed for advanced human interactions and is used as the default model over GPT-3.5 due to its cost-effectiveness and superior performance.
Learn More: All You Need to Know About ChatGPT
GPT-3.5 vs GPT-4: A Comparison
OpenAI announced its latest language model, GPT-4, on March 14, 2023.
They have not disclosed the exact number of parameters they used to train GPT-4. However, estimates suggest that it’s about ten times greater than that of GPT-3.5, indicating a substantial leap in complexity and processing power.
Learn more about GPT-3.5 vs GPT-4 comparison.
A key distinguishing feature between GPT-4 and its predecessor is its multimodal capability. GPT-4 has the ability to accept both text and image inputs, making it a significantly more versatile model. In contrast, GPT-3.5 only supports text inputs.
Learn More: GPT-4 Image Input How to Use it?
The character limit for the most advanced version of GPT-4 is approximately 25,000, double that of GPT-3.5, which sits around 12,000. This allows GPT-4 to handle much larger chunks of information at once, potentially improving comprehension and response quality.
In terms of cognitive performance, GPT-4 is designed with enhanced reasoning abilities and creativity, enabling it to perform better in complex problem-solving scenarios. In tests, GPT-4 outperformed GPT-3.5, scoring higher in approximate percentiles among test-takers. Learn the performance comparisons in the exam scores between GPT-3.5 and GPT-4.
However, these enhancements come with increased costs and slower response times. The usage fee for the GPT-4 API is about 20 times higher than GPT-3.5.
Additionally, GPT-4 responds about 1.7 times slower than GPT-3.5, likely due to the increased computational load from the larger parameter count and multimodal capabilities.