We’re now into the third year of the AI boom, and industry leaders are showing no signs of slowing down, pushing out newer and (presumably) more capable models on a regular basis. ChatGPT, of course, remains the undisputed leader.
But with more than a half-dozen models available from OpenAI alone, figuring out which one to use for your specific project can be a daunting task.
o1
OpenAI CEO Sam Altman spent a better part of his summer in 2024 teasing the imminent release of the company’s secretive Project Strawberry. The company finally unveiled the new system in September, outing it as OpenAI’s first “reasoning” model and renaming it “o1.” Much like the two-stage release of GPT-2, where a stripped-down version was released months ahead of the full-feature version, GPT Plus subscribers were first given access to the o1-preview version.
As a reasoning model, o1 is touted as using humanlike reasoning to check its own answers before outputting more accurate responses to complicated science, math, and coding queries. Trained on a “completely new optimization algorithm and a new training dataset specifically tailored for it,” per the company, o1 drastically outperforms the GPT-4o family of models across industry benchmark tests.
The full version of OpenAI o1 made its debut in December 2024 during the company’s 12 Days of OpenAI livestream event. Its internal system of checks reportedly helps the new model hallucinate less than its predecessors. According to OpenAI’s tests, the new version of o1 reduces “major errors” on “difficult real-world questions” by 34% compared to its preview version. If you’ve got reams of highly complex data to pour over or have complicated math and coding problems to solve, o1 is the OpenAI model you’ll want to use.
Unfortunately, it is not currently available to free-tier users. You can access o1 with a Plus or Teams subscription, though you’ll only get 50 messages with it per week at that level. You’ll need to shell out $200 monthly for a Pro subscription if you want to use it without limits. So, our recommendation would be to start out on 01, and if you run out of your allotted messages for the day, move on to one of the models below.
o1-mini
OpenAI’s o1-mini is a lightweight version of the company’s larger reasoning model. It is geared more toward quickly solving challenging mathematics and coding problems rather than creative prose generation. It’s built for speed and designed to provide fast, efficient solutions for specific technical tasks, making it ideal for task automation.
Unlike the preview model, you can still access o1-mini, though you will need at least a Plus subscription. With Plus, you’ll get 50 messages a day with OpenAI o1-mini. Pro subscribers, aka the folks shelling out $200 a month, have unlimited access to the mini model.
03
o1 enjoyed its position at the peak of OpenAI’s model mountain for less than a fortnight. On the twelfth and final livestream of the company’s 12 Days of OpenAI event, CEO Sam Altman revealed the o1’s successor model, dubbed o3 to avoid a copyright battle with a UK telecom.
o3, our latest reasoning model, is a breakthrough, with a step function improvement on our hardest benchmarks. we are starting safety testing & red teaming now. https://t.co/4XlK1iHxFK
— Greg Brockman (@gdb) December 20, 2024
Billed as a “breakthrough, with a step function improvement on our hardest benchmarks,” by OpenAI President and Co-Founder, Greg Brockman, the o3 model shares the same reasoning architecture as the o1 family. It just does it better. The o3 model enables users to select between low, medium, and high compute resource allocations, with high compute returning the most complete results, albeit at the cost of inference time and financial cost. ARC-AGI co-creator Francois Chollet has noted that high compute reportedly will cost thousands of dollars per task.
The o3’s performance is nothing to sneeze at. According to OpenAI, o3 beats out the o1’s performance by nearly 23 percentage points on the SWE-Bench Verified coding test, more than 60 points higher on the Codeforce benchmark, and missed just one question on the AIME 2024 mathematics test. It also reportedly outperformed human experts on the GPQA Diamond test and correctly answered a quarter of the questions on the near impossible EpochAI Frontier Math benchmark where conventional LLMs barely get 2% of them right.
The o3 model is currently available in early access to safety and security researchers. It’s expected to make its public debut in early 2025.
o3-mini
OpenAI’s o3-mini model serves the same purpose as o1-mini or GPT-4o-mini: to serve as a lightweight and more efficient version of the larger full model. Announced in late January, o3-mini is expected to be released as both an API and as an available ChatGPT model “in a couple weeks,” Sam Altman teased in a post on X.
In a marked departure for the company, Altman also confirmed that o3-mini will become the baseline model for its chatbot platform — and therefore free to use, regardless of subscription tier — upon its release. Typically the company has offered an older generation model for free-tier users along with limited access to the newer and better ones, as a means of encouraging subscriber sign-ups.
Interestingly, existing paid tier subscribers use GPT-4o Turbo (the company’s current flagship conventional model) as the basis for their ChatGPT conversations. There’s no word yet on whether GPT-4o Turbo will remain the baseline model for paid subscribers, given Altman’s promise of “tons of o3-mini usage.”
OpenAI publicly released o3-mini on January 31st, 2025. Free tier users get limited access to the model, while Plus and Teams subscribers can make up to 150 queries per day, and Pro subscribers receive unlimited access to it and all of OpenAI’s various models.
“While o1 remains our broader general-knowledge reasoning model, o3-mini provides a specialized alternative for technical domains requiring precision and speed,” the company wrote in its announcement post. “The release of o3-mini marks another step in OpenAI’s mission to push the boundaries of cost-effective intelligence.”
GPT-4o
GPT-4o is to GPT-4 in terms of performance what GPT-4 was to GPT-3.5. Introduced in May 2024, GPT-4o (the “o” stands for “omni”) is OpenAI’s “latest, fastest, highest intelligence model.” It offers human-level response times (critical for features like Advanced Voice Mode), improved performance with languages beyond English, and a far better understanding of vision and audio content.
Not only is this model faster, more efficient, and more cost-effective than its predecessors, but 4o has also achieved state-of-the-art results in multilingual and vision benchmark tests. It also offers a 128,000-character context window and can generate text, images, audio, and computer code. It recently received a minor boost to its creative writing skills thanks to a pre-Thanksgiving update.
You can access GPT-4o through the free tier of ChatGPT, but you will run into restrictive usage caps. Once you reach your limit with 4o queries, the platform will drop your access down to 4o-mini. Subscribing to the $20-per-month Plus tier will significantly increase that usage cap.
GPT-4o mini
OpenAI’s “lightest-weight intelligence model” is GPT-4o mini. Designed more for completing computationally simple tasks performed quickly and repeatedly as opposed to pondering over challenging coding or analytical problems, 4o mini offers nearly all of the same features as the larger 4o model, save for its restricted access to some of the more advanced analytical tools.
GPT-4o mini is available to Plus and Pro subscribers, and you can select it from the drop-down model list in the upper left corner of the ChatGPT home screen. You can also access it as a free-tier user by blowing through your GPT-40 usage allowance. The platform will shunt you over to the 4o mini model until you come off cool down and are allowed to access GPT-4o again.
GPT-4
GPT-4 was the first big step up in ChatGPT’s capabilities since its debut. Launched in April 2023 alongside the new ChatGPT Plus subscription tier, GPT-4 powered both ChatGPT and Microsoft’s free Copilot platform. OpenAI touted GPT-4 as “more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5.” It was also the company’s first multimodal model, allowing it to generate and analyze images and audio in addition to text.
GPT-4 still serves as the base model available for free-tier ChatGPT users. For most casual users, GPT-4 and its “omni” variants are plenty capable of performing the inference tasks that you need.
Early versions
GPT-1, the original large language model from OpenAI, made its debut in June 2018 and became the archetype of generative pre-trained transformer technology. Even in its earliest form, GPT-1 managed to outperform the state-of-the-art models of the time by nearly 6% at natural language inference tasks.
Its 1.5-billion-parameter successor, GPT-2, arrived in November 2019. It expanded upon GPT-1’s basic text generation capabilities to include answering questions, summarizing documents, and translating text between languages. GPT-3 arrived in May 2020 but was quickly snatched up by Microsoft that September with a licensing scheme that gives Microsoft exclusive use of that model.
The GPT-3.5 model is a subset of Microsoft’s licensed GPT-3 that offered better performance and a later knowledge cutoff date than earlier iterations. When OpenAI launched the ChatGPT platform in November 2022, the chatbot ran atop the GPT-3.5 model. The company incorporated web browsing capabilities the following April, but by May 2023, GPT-3.5 had been depreciated in favor of the far more capable GPT-4 family of models. As such, it’s no longer available for public use and can only be accessed through OpenAI’s developer API.
