OpenAI Unveils New Model Behind ChatGPT
While OpenAI is best known for creating ChatGPT, I’ve been focusing primarily on their work with AI video generation. Sora , the company’s short AI video generator, is already tricking social media users with hyper-realistic AI effects. I fear we’ve reached a point where it’s too difficult to believe what you see in your feeds is real.
But today, we’ll take a break from video AI and return to OpenAI’s flagship product. On Wednesday, the company announced the latest versions of GPT , the model that powers ChatGPT: GPT-5.1 Instant and GPT-5.1 Thinking. OpenAI says Instant is the company’s most widely used model, but it’s “now faster, smarter, and better at following your instructions.” Thinking, on the other hand, is the company’s reasoning model, which is faster and “easier to understand” when performing simple tasks, and more “persistent” when executing complex queries.
I think these are mostly minor updates, but here’s what’s new, according to OpenAI:
GPT-5.1 Instant
Let’s first look at Instant. OpenAI claims the model is “surprisingly” playful. To the question “I’m feeling stressed and could use some relaxation tips,” OpenAI shows two example responses for GPT-5 and GPT-5.1 Instant, respectively:
-
GPT-5: “Here are some simple and effective ways to relieve stress—you can combine them depending on how you feel and how much time you have.”
-
GPT-5.1 Instant: “I understand, Ron—it’s completely normal, especially given everything that’s been going on lately. Here are some ways to relax, depending on how stressed you are.”
Sure, I see the differences: GPT-5 sounds a bit more serious, while 5.1 takes a classic, overly friendly AI approach and always includes the username. How human . But scrolling through the rest of the comparison, I’m not sure I see that much of a difference. Both options break their sentences into bullet points, though 5.1 sometimes uses more flowery language, like “bonus points.”
Another change OpenAI touts is improved instruction following. In the examples provided, a user asks the bot to “Always answer in six words.” GPT-5 follows suit in its first response, but when asked where the user should travel this summer, the bot completely ignores the six-word requirement and offers a full answer. GPT-5.1, on the other hand, maintains this dynamic throughout the entire conversation, regardless of the complexity of the questions. I realize this is just an example, but why would anyone want a bot to follow such a strange request? I assume this implies that GPT-5.1 will preserve your original instructions throughout the entire interaction, but if so, why not show an example where this is actually useful? I’d rather get a more complete answer to important questions than have all answers consist of exactly six words.
Finally, OpenAI claims that Instant can still use reasoning models when needed, while still delivering results quickly. The company doesn’t elaborate on this, but they must have made some changes to the algorithm that allow the model to “think through” complex sentences without delving into too much detail to the point where it takes too long to respond.
GPT-5.1 Thinking
GPT-5.1 Thinking is a direct upgrade to the existing GPT-5 Thinking model. Like GPT-5.1 Instant, OpenAI claims that this model can adapt its processing to the task at hand. This theoretically means that 5.1 Thinking can respond faster to simpler cues while spending more time “thinking” through complex tasks. OpenAI claims that GPT-5.1 Thinking responds approximately twice as fast to the “fastest tasks” and half as slow to the “slowest tasks” compared to GPT-5 when both models are configured to “Standard.”
OpenAI claims that 5.1 Thinking uses less technical language and doesn’t necessarily assume you understand complex terms. The company shows an example of someone asking the bot to explain BABIP and wRC+ (batting average per game and weighted runs batted in plus, respectively). GPT-5’s explanation includes abbreviations, formulas, and insider concepts that I certainly wouldn’t have grasped as someone who knows next to nothing about baseball. On the other hand, 5.1 Thinking’s output deciphers these abbreviations, walks the user through the formulas rather than simply displaying them, and explores some complex topics in more detail. I can see how this could improve clarity, although one should always be wary of hallucinations . While it’s great to make answers more understandable, it won’t help if the answers themselves are completely inaccurate.
ChatGPT tone setting
As part of these new models, OpenAI is changing the way ChatGPT customizes tone and style. The new personalization options include the familiar “Default” tone, but “Listener” is now called “Friendly” and “Robot” is now called “Efficient.” (Perhaps GPT took offense to the latter.) OpenAI says all of these options have been updated, but doesn’t specify which ones. You also have new options to choose from, including “Professional” (subtle and precise), “Frank” (direct and encouraging), and “Quirky” (playful and inventive).
Finally, the company says it’s currently testing a new settings option that lets you customize the bot’s tone and style, but is only launching the experiment with a limited group of users.
What else is new?
OpenAI states that by default, you won’t have to choose between these two models: GPT-5.1 Auto will determine which model is best based on your query. This has been the case since GPT-5’s launch , although paid subscribers can still choose the model manually.
Regarding paid subscribers, GPT-5.1 will initially be available to Pro, Plus, Go, and Business users. The company began shipping the model on Wednesday and promises to gradually roll it out over the next few days. The model will soon be available to free users and those not logged in, although the exact timing has not yet been announced.
Paid subscribers will also have access to GPT-5 for three months, as OpenAI is in no rush to phase out support for the older version. The company likely doesn’t want to repeat the problems that arose when models like GPT-4o and earlier were immediately removed from ChatGPT, as users who were deeply attached to this model were upset by its loss. OpenAI notes that deprecating GPT-5 will not affect existing legacy models, such as 4o, that are still available.
For a new GPT update, this isn’t all that impressive. Perhaps that’s for the best: OpenAI overestimated GPT-5, which disappointed many GPT-4 fans, especially when the company subsequently abandoned that model. Keeping expectations for GPT-5.1 in check is probably a smart move, but I can’t help but wonder if this is a sign of AI slowing down. We know that AI companies are running out of training data for their models, we just don’t know how quickly this will manifest itself. GPT-5.1 probably doesn’t herald doom for OpenAI, but interestingly, it doesn’t represent an exponential improvement either.
Disclosure: Lifehacker’s parent company, Ziff Davis, filed a lawsuit against OpenAI in April, alleging that it infringed Ziff Davis’s copyrights in the training and operation of its artificial intelligence systems.