ChatGPT: Which model to choose?

We are now in the third year of the AI ​​boom and industry leaders show no signs of slowing down, regularly launching new, more capable models. ChatGPT, unsurprisingly, remains the undisputed leader.

However, with over half a dozen models available from OpenAI alone, choosing the best one for your specific project can be a tricky task.

GPT o1

If you want to test out ChatGPT’s most advanced model, look no further than o1.

Sam Altman, the CEO of OpenAI, spent much of his summer 2024 revealing the impending release of his secret project, Project Strawberry. This project was finally presented in September under the name “o1”, OpenAI’s first reasoning model. Like the two-stage unveil of GPT-2, the preview version of o1 was first available to GPT Plus subscribers.

As a reasoning model, o1 uses human-like reasoning to verify its own answers before providing more accurate results to complex questions in science, mathematics, and programming. According to OpenAI, this model significantly outperforms the GPT-4o family of models in industry benchmark tests.

The full version of OpenAI o1 was launched in December 2024 during the “12 Days with OpenAI” livestream event. Its internal verification system would help reduce the hallucinations experienced with previous models. According to OpenAI testing, this version of o1 reduces “major errors” in “difficult real-world questions” by 34% compared to its preview version. If you need to process particularly complex data or solve difficult math and programming problems, o1 is the OpenAI model to choose.

Unfortunately, it’s not yet available to free version users (and probably won’t be for a while). You can access it with a Plus or Teams subscription, but you’ll be limited to 50 messages per week with this subscription level. If you want to use it without limits, a Pro subscription at $200 per month will be necessary. It is therefore recommended to start with o1, and if you reach your daily quota, to turn to one of the models below.

GPT o1-mini

The o1-mini version of OpenAI is a lightweight version of the larger reasoning model. It is designed primarily for quickly solving math and programming problems rather than generating creative text. Its goal is to offer effective solutions for specific technical tasks, making it ideal for task automation.

Unlike the preview version, o1-mini is still accessible, although at least a Plus type subscription is required. With this plan, you are entitled to 50 messages per day with OpenAI o1-mini. Pro subscribers, who pay $200 per month, have unlimited access to the mini model.

GPT-4o

GPT-4o represents the performance evolution over GPT-4, just as GPT-4 was over GPT-3.5. Introduced in May 2024, GPT-4o (the “o” standing for “omni”) is touted as OpenAI’s fastest and smartest model to date. It enables human-like response times, which is crucial for features like advanced voice mode, while improving performance in languages ​​other than English and better understanding of visual and audio content.

Not only is this model faster, more efficient and more cost-effective than its predecessors, but it also achieved industry-leading results in multilingual and visual benchmark tests. With a context of 128,000 characters, it can generate text, images, audio and computer code. Her creative writing skill recently got a slight boost thanks to an update before Thanksgiving.

You can access GPT-4o through the free version of ChatGPT, but you will face restrictive usage limits. Once you reach your quota with 4o, the platform will redirect you to either the 4o-mini model or the older GPT-4-Turbo model. Subscribing to the Plus tier at $20 per month will significantly increase this usage quota.

GPT-4o mini

OpenAI’s “lightest intelligence” model is GPT-4o mini. Designed to quickly perform simple tasks rather than tackle complex analytical problems, 4o mini offers almost all of the same features as the larger 4o model, with the exception of restricted access to some advanced analytical tools.

GPT-4o mini is available to Plus and Pro subscribers, and you can select it from the drop-down menu at the top left of the ChatGPT home screen. Free version users can also access it, but only after exhausting their GPT-4o usage quota. The platform will then transfer you to the 4o mini model until you can access GPT-4o again.

GPT-4

GPT-4 was the first major advancement in ChatGPT’s capabilities since its launch. Unveiled in April 2023, with the new ChatGPT Plus subscription tier, GPT-4 powered both ChatGPT and Microsoft’s free Copilot platform. OpenAI touted GPT-4 as being “more reliable, creative, and capable of handling much more nuanced instructions than GPT-3.5.” It was also the company’s first multimodal model, capable of generating and analyzing images and audio in addition to text.

GPT-4 remains the base model offered to free ChatGPT users. For most casual users, GPT-4 and its “omni” variants are largely capable of performing the necessary inference tasks.

Previous versions

GPT-1, OpenAI’s first large language model, was launched in June 2018 and has become the archetypal model of generative pre-trained transformer technology. Even in its most primitive form, GPT-1 managed to outperform state-of-the-art models of the time by 6% in natural language inference tasks.

Its successor, GPT-2, with its 1.5 billion parameters, arrived in November 2019. It significantly expanded GPT-1’s text generation capabilities to include question answers, document summaries, and translations between languages. GPT-3 first appeared in May 2020 but was quickly acquired by Microsoft in September, with a licensing agreement that grants Microsoft exclusive use of this model.

The GPT-3.5 model is a subcategory of GPT-3 licensed by Microsoft, offering better performance and a more recent cutoff date than its previous iterations. When OpenAI launched the ChatGPT platform in November 2022, the chatbot ran on the GPT-3.5 model. The company integrated web browsing capabilities again the following April, but as of May 2023, GPT-3.5 was deprecated in favor of the much more capable GPT-4 family of models. Therefore, it is no longer accessible to the public and can only be used via the OpenAI Developer API.