- AI
- A
Getting the Most Out of ChatGPT: Tips for Choosing the Right Model
Hello! This is Andrey from the “Runaway Neural Network” channel, where I talk about AI from a creative perspective. In my work, I use models from various companies, including, of course, ChatGPT with a Plus subscription. It’s one of the best AIs on the market, but it has one big problem—the range of available models is so vast that even experienced users can get confused. Below, I’ll explain which model I use and in which cases.
ChatGPT-4o
A classic chatbot that I use when I want answers to not-too-complicated questions, to write fiction (a story, a fairy tale for my daughter), or just to chat. Outside the US, 4o is the only model that gathers knowledge about you based on previous chats, meaning it gives more personalized answers. For example, you can try a simple prompt — “Based on what you know about me, suggest some new topics for conversation.” At the moment, ChatGPT-4o in the Plus subscription can be used virtually without limits, so feel free to experiment.
ChatGPT o3
Currently, this is the most powerful model in the subscription, which I use for complex tasks: analyzing data, fact-checking a text, taking an idea and evaluating its strengths and weaknesses, as well as pointing out gaps. ChatGPT o3 also performs web searches across several dozen websites and provides a structured response.
At the same time, I recommend treating ChatGPT o3 not as a chatbot, but as a report generator. First, the model is analytical, which means it can take up to a couple of minutes to write an answer. Secondly, o3 simply loves lists and tables. Finally, as part of the Plus subscription, this model is limited to 200 requests per week—this is enough if you provide the model with a clear initial prompt, giving as much information as possible and explaining what you want to see in the answer. My advice—if you're not sure about your request, ask the model not to answer right away, but to ask you additional questions that will help it formulate the final response. This will take an extra request, but the likelihood of the model making a mistake is much lower.
Codex-1
This is a coding agent based on the same ChatGPT o3, but additionally trained on programming. Codex-1 can connect to your repository, perform several tasks in parallel, run tests, and so on. Most importantly—unlike the regular o3, the use of Codex-1 is currently not limited. So if you want to try ChatGPT for programming, this is where you should start. Keep in mind that Codex is selected from the left menu, not from the model selection menu.
ChatGPT o4-mini-high
o4 is a completely new generation of OpenAI models, trained after ChatGPT o3. However, the "mini" suffix means it’s a lightweight version, seriously limited in terms of knowledge. But it works faster and allows up to 100 requests per day (and 300 requests for the even more simplified o3-mini, but I don’t use it).
I personally learned to choose between o3 and o4-mini-high intuitively, but if you’re just starting to use ChatGPT Plus, my advice is—try making 20-30 requests with o3, simultaneously “feeding” them into o4-mini-high and comparing the results.
Also, ChatGPT o4-mini-high writes better in Russian than o3, so when the latter gives answers that are too convoluted, I switch to o4-mini-high and ask it to explain the last message in simpler language.
DeepResearch
A special mode of the ChatGPT o3 model for deep web search – the model reads your task, asks additional questions, and returns in 15–30 minutes with a very detailed report. DeepResearch is limited to 25 requests per month, and after 10 requests, the function switches to the o4-mini-high model, which worsens the answer quality. However, after the launch of the “regular” o3, I almost stopped using DeepResearch — it's much more efficient to throw the same request into o3, then, if needed, make several additional queries and, in the end, ask for a summary of the entire thread.
ChatGPT-4.1
A fast model for coding, whose main advantage is its huge context window — 1 million tokens is enough to easily fit “War and Peace.” Unfortunately, on the Plus subscription, the context window is much smaller — only 32 thousand tokens — which makes this model not very appealing. You can also select the GPT-4.1-mini model, which is a bit simpler and a bit faster.
ChatGPT-4.5
I consider this model one of OpenAI's biggest flops in recent years. It was developed for a long time under the code name Project Orion and was seen as the chatbot meant to replace ChatGPT-4o. At launch, ChatGPT-4.5 indeed gave higher quality answers than 4o — especially for creative tasks like writing and humor. But that came at a steep price: the model turned out to be large, expensive, and slow, which is critical for a chatbot. Because of its cost, ChatGPT-4.5 was initially limited to 50 requests per week on the Plus subscription, and now the limit has even dropped to 10 requests, making it nearly unusable. Furthermore, OpenAI has recently released several updates to ChatGPT-4o, after which its answer quality, in my opinion, has even surpassed 4.5.
The verdict on ChatGPT-4.5 is simple — play with its creative possibilities by asking it to write something in a chosen genre or on a chosen topic. And then leave it behind without a second thought — I'm pretty sure the model will simply be discontinued after ChatGPT-5 is released.
There's a chance that the torturous necessity of choosing a model for responses in ChatGPT will end in July — rumors say that's when ChatGPT-5 will be released, which will “decide” on its own how to handle user requests. In the meantime, I hope my experience will be useful to you.
P.S. You can traditionally support me by subscribing to my Telegram channel “runaway neural network”, where I talk about AI from a creative perspective.
Write comment