GPT-3.5 vs. GPT-4: Biggest differences to consider
GPT-3.5 or GPT-4? With multiple OpenAI language models to choose from, picking the right option for your organization's needs comes down to the details.
With a growing number of underlying model options for OpenAI's ChatGPT, choosing the right one is a necessary first step for any AI project. Knowing the differences between GPT-3, GPT-3.5 and GPT-4 is essential when purchasing SaaS-based generative AI tools.
GPT-3.5, the refined version of GPT-3 rolled out in November 2022, is currently offered both in the free web app version of ChatGPT and via the paid Turbo API. GPT-4, released in March 2023, offers another GPT choice for workplace tasks. It powers ChatGPT Team and ChatGPT Enterprise, OpenAI's first formal commercial enterprise offerings. GPT-4 also entails additional features like multimodality and API implementation considerations.
Choosing between GPT-3.5 and GPT-4 means parsing out the differences in their respective features. By breaking down the two models' key differences in capabilities, accuracy and pricing, organizations can decide which OpenAI GPT model is right for them.
GPT-3.5 vs. GPT-4: The major differences
GPT-3.5 and GPT-4 are both versions of OpenAI's generative pre-trained transformer model, which powers the ChatGPT app. They're currently available to the public at a range of capabilities, features and price points.
This article is part of
What is Gen AI? Generative AI explained
Extended capabilities
The difference in capabilities between GPT-3.5 and GPT-4 indicates OpenAI's interest in advancing their models' features to meet increasingly complex use cases across industries.
GPT-3.5
GPT-3.5 has several key capabilities:
- Understand and generate human-like text using natural language comprehension and generation to complete various natural language-related tasks.
- Translate text from one language to another with some fluency and accuracy.
- Answer questions by providing relevant information, making it suitable for chatbots and virtual assistants using GPT-3.5 Turbo, which is tailored to working with the Chat Completions API.
- Generate concise summaries of longer text, such as documentation and reports.
- Generate content for a wide range of use cases and writing projects, such as emails and code.
Additionally, developers can fine-tune GPT-3.5 Turbo on their data, tailoring the model to specific use cases and enhancing its performance on tasks. Users can access two model variants through the GPT-3.5 Turbo API:
- Gpt-3.5-turbo-0125 supports a 16,385-token context window and is optimized for dialogue with higher accuracy than previous GPT-3.5 Turbo models.
- Gpt-3.5-turbo-instruct is an instruction model that only supports a 4,096-token context window.
GPT-3 vs. GPT-3.5
In June 2020, OpenAI released GPT-3. Following GPT-1 and GPT-2, the vendor's previous iterations of generative pre-trained transformers, GPT-3 was the largest and most advanced language model yet. As a large language model, it works by training on large volumes of internet data to understand text input and generate text content in a variety of forms.
In November 2022, OpenAI released its chatbot ChatGPT, powered by the underlying model GPT-3.5, an updated iteration of GPT-3. While sometimes still referred to as GPT-3, it is really GPT-3.5 that is in use today.
GPT-4
OpenAI designed GPT-4 to be more reliable, creative and capable of handling nuanced instructions than its predecessors. GPT-4's extended capabilities include the following:
- Multimodality. GPT-4 can accept images as inputs to generate captions, classifications and analyses.
- Larger context windows. GPT-4 offers two context window sizes -- 8,192 and 32,768 tokens -- with which the model can handle over 25,000 words of text. This enables use cases like long-form content creation, extended conversations, and document search and analysis.
- Broader general knowledge. GPT-4's expanded knowledge base enables the model to generate, edit and iterate various tasks such as composing songs, writing screenplays or learning a user's writing style.
- Enhanced safety and alignment features. In response to market concerns, OpenAI undertook efforts to make GPT-4 safer and more aligned with user intentions. According to OpenAI's internal evaluations published in April 2023, GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5.
The GPT-4 API includes the Chat Completions API (97% of GPT API usage as of July 2023). It supports text summarization in a maximum of 10 words and even programming code completion. Chat Completions API also provides few-shot learning capabilities. OpenAI plans to focus more attention and resources on the Chat Completions API and deprecate older versions of the Completions API.
Accuracy
GPT-3.5 is only trained on content up to September 2021, limiting its accuracy on queries related to more recent events. GPT-4, however, can browse the internet and is trained on data up through April 2023 or December 2023, depending on the model version.
Unfortunately, Stanford and University of California, Berkeley researchers released a paper in October 2023 stating that both GPT-3.5 and GPT-4's performance has deteriorated over time. In line with larger conversations about the possible issues with large language models, the study highlights the variability in the accuracy of GPT models -- both GPT-3.5 and GPT-4.
Availability and pricing
As vendors start releasing multiple versions of their tools and more AI startups join the market, pricing will increasingly become an important factor in AI models. To implement GPT-3.5 or GPT-4, individuals have a range of pricing options to consider.
GPT-3.5
GPT-3.5 is currently available in the free version of ChatGPT.
The following table details GPT-3.5 Turbo API costs:
| GPT-3.5 Turbo API pricing | ||
| Model | Input | Output |
| gpt-3.5-turbo-0125 | $0.0005 / 1K tokens | $0.0015 / 1K tokens |
| gpt-3.5-turbo-instruct | $0.0015 / 1K tokens | $0.0020 / 1K tokens |
GPT-4
GPT-4 is available on ChatGPT Plus for $20 per month per person. It's also available as ChatGPT Team, which costs $25 per person per month, and as ChatGPT Enterprise, which requires prospective buyers to contact OpenAI's sales team for pricing.
The following table details GPT-4 API costs:
| GPT-4 API pricing | ||
| Model | Input | Output |
| gpt-4 | $0.03 / 1K tokens | $0.06 / 1K tokens |
| gpt-4-32k | $0.06 / 1K tokens | $0.12 / 1K tokens |
Introduction to GPT-4 Turbo
The new GPT-4 Turbo is the latest generation model from OpenAI. OpenAI API accounts with GPT-4 access can access GPT-4 Turbo under the model name gpt-4-1106-preview. Currently, the GPT-4 Turbo API is available at a lower price than GPT-4.
OpenAI describes GPT-4 Turbo as more powerful than GPT-4, and the model is trained on data through December 2023. It has a 128,000-token context window, equivalent to sending around 300 pages of text in a single prompt. It's also three times cheaper for input tokens and two times more affordable for output tokens than GPT-4, with a maximum of 4,096 output tokens.
| GPT-4 Turbo API pricing | ||
| Model | Input | Output |
| gpt-4-0125-preview | $0.01 / 1K tokens | $0.03 / 1K tokens |
| gpt-4-1106-preview | $0.01 / 1K tokens | $0.03 / 1K tokens |
| gpt-4-1106-vision-preview | $0.01 / 1K tokens | $0.03 / 1K tokens |
The GPT-4 Turbo API's rate limits depend on the usage tier. Usage tier information can be found in a user's Limits settings page. Since this model is a preview, OpenAI won't currently accommodate rate limit increases on GPT-4 Turbo. OpenAI plans to release a stable, general availability GPT-4 Turbo model, but they've yet to announce a release date.
Will Kelly is a technology writer, content strategist and marketer. He has written extensively about the cloud, DevOps and enterprise mobility for industry publications and corporate clients and worked on teams introducing DevOps and cloud computing into commercial and public sector enterprises.
