GPT-5 is out! What makes it different?

Did you notice that all of your model options in ChatGPT suddenly went away?

Previously, we had lots of model options, described with a perhaps less than helpful naming convention.

Yesterday, OpenAI, the makers of ChatGPT, announced their latest model. They had promised a “unifying” model, and they really went all in. Instead of selecting a model before you enter a prompt and trying to remember the difference (wait, is 4o an o-series model, or more like GPT4?? what does “mini-high” mean?) you may now (or will soon) have just two options: GPT-5 and GPT-5 Thinking.

Perhaps you, like me, had long felt that all of these confusingly-named models ought to be simplified, and then when you suddenly didn’t have them, you wondered whether you were missing something?

I had to borrow this image because mine STILL hasn’t updated!

GPT-5 is a new model, and it’s also the result of an emerging new approach to chat bots.

Instead of selecting the model yourself, when you type a prompt into GPT-5, it automatically routes your request. Is your prompt best served by a multi-step analytic process with dependencies (a request previously best served by an o-series model?) It will send your prompt to a model that designed to do that—one structured and optimized to do “chain-of-thought.” (OpenAI sometimes calls this “reasoning,” but I do not.) If it is a more straightforward request that you might have used a 4-series model for, it will treat it appropriately, skipping the resource-intensive and time-consuming process that an o-series model would have done and sending it to a model designed to respond quickly. If you need an image, it will send your request to a specialized model for that, too.

You do have one other option: GPT-5 Thinking. “Thinking” is another word OpenAI uses for the chain-of-thought models. I suspect what “GPT-5 Thinking” does is reduce the options for routing and give your prompt directly to its chain-of-thought model.

Is it better & how?

Open AI claims that GPT-5 performs better on the types of problems that previous models were not great at. For example, it allegedly has better emotional intelligence, writing skills, and math ability. They report that it hallucinates less often than GPT4o, although different benchmarks show different degrees of improvement: it likely varies across domains and circumstances, so we will have to wait for the full picture to emerge. So far, sycophancy seems to be less of a problem as well, or at least it’s easier to nudge the model away from enthusiastically praising your every idea.

What is clear:
GPT 5 has a much bigger context window, which was a major downside of previous ChatGPT models compared to Claude. This means that it can take in a much larger amount of information than before: it can now process a fairly long book in its entirety (and write a short book back). The total context window is 400k tokens, or about 300k words (a normal, non-fiction book is 50-70k, and A Song of Ice and Fire, the first Game of Thrones book, was just under 300k words).

It is more widely available. A limited version of GPT-5 is available for free users, breaking OpenAI’s habit of giving free users access to past generation models.

People who use it for coding like it. My friends who are doing more sophisticated coding projects than me are thrilled with GPT-5’s coding ability. I trust them, but that’s as much assurance as I can give ya for now :)

People who used 4o for emotional support do not. The ChatGPT subreddit is full of people who are referring to the loss of 4o as “mentally devastating,” an “irreversible emotional loss,” and saying “it feels like someone died.” It is possible that the models they had relied on were finely tuned to their preferences by many many chats, and GPT-5 will be able to replicate that, but for now, these users find the change very disruptive.

I will take this opportunity to encourage readers to be very cautious about relying on any AI model for emotional support.

I look forward to spending time in GPT-5 and of course I will update you if I learn anything interesting or important!

—
LLM disclosure.
Nah, I just wrote this one :)