OpenAI Releases GPT-4.1: Why This Super-Powered AI Model Will Kill GPT-4.5

OpenAI unveiled GPT-4.1 on Monday, a trio of new AI models with context windows of up to one million tokens—enough to process entire codebases or small novels in one go. The lineup includes standard GPT-4.1, Mini, and Nano variants, all targeting developers.

The company’s latest offering comes just weeks after releasing GPT-4.5, creating a timeline that makes about as much sense as the release order of the Star Wars movies. “The decision to name these 4.1 was intentional. I mean, it’s not just that we’re bad at naming,” OpenAI product lead Kevin Weil said during the announcement—but we are still trying to find out what those intentions were.

GPT-4.1 shows pretty interesting capabilities. According to OpenAI, it achieved 55% accuracy on the SWEBench coding benchmark (up from GPT-4o’s 33%) while costing 26% less. The new Nano variant, billed as the company’s “smallest, fastest, cheapest model ever,” runs at just 12 cents per million tokens.

Also, OpenAI won’t upcharge for processing massive documents and actually using the one million token context. “There is no pricing bump for long context,” Kevin emphasized.

The new models show impressive performance improvements. In a live demonstration, GPT-4.1 generated a complete web application that could analyze a 450,000-token NASA server log file from 1995. openAI claims the model passes this test with nearly 100% accuracy even at million tokens of context.

Michelle, OpenAI’s post-training research lead, also showcased the models’ enhanced instruction-following abilities. “The model follows all your instructions to the tea,” she said, as GPT-4.1 dutifully adhered to complex formatting requirements without the usual AI tendency to “creatively interpret” directions.

How Not to Count: OpenAI’s Guide to Naming Models

The release of GPT-4.1 after GPT-4.5 feels like watching someone count “5, 6, 4, 7” with a straight face. It’s the latest chapter in OpenAI’s bizarre versioning saga.

After releasing GPT-4 it upgraded the model with multimodal capabilities. The company decided to call that new model GPT-4o (“o” for “omni”), a name that could be also be read as “four zero” depending on the font you use

Then, OpenAI introduced a reasoning-focused model that was just called “o.” But don’t confuse OpenAI’s GPT-4o with OpenAI’s o because they are not the same. Nobody knows why they picked this name, but as a general rule of thumb, GPT-4o was a “normal” LLM whereas OpenAI o1 was a reasoning model.

A few months after the release of OpenAI o1, came OpenAI o3.

But what about o2?—Well, that model never existed.

“You would think logically (our new model) maybe should have been called o2, but out of respect to our friends at Telefonica—and in the grand tradition of open AI being really truly bad at names—it’s going to be called o3,” Sam Altman said during the model’s announcement.

The lineup further fragments with variants like the normal o3 and a smaller more efficient version called o3 mini. However, they also released a model named “OpenAI o3 mini-high” which puts two absolute antonyms next to each other because AI can do miraculous things.In essence, OpenAI o3 mini-high is a more powerful version than o3 mini, but not as powerful as OpenAI o3—which is referenced in a single chart by Openai as “o3 (Medium),” as it should be. Right now ChatGPT users can select either OpenAI o3 mini or OpenAI o3 mini high. The normal version is nowhere to be found.

Also, we don’t want to confuse you anymore, but OpenAI already announced plans to release o4 soon. But, of course, don’t confuse o4 with 4o because they are absolutely not the same: o4 reasons—4o does not.

Now, let’s go back to the newly announced GPT-4.1. The model is so good, it is going to kill GPT-4.5 soon, making that model the shortest living LLM in the history of ChatGPT. “We’re announcing that we’re going to be deprecating GPT-4.5 in the API,” Kevin declared, giving developers a three-month deadline to switch. “We really do need those GPUs back,” he added, confirming that even OpenAI can’t escape the silicon shortage that’s plaguing the industry.

At this rate, we’re bound to see GPT-π or GPT-4.√2 before the year ends—but hey, at least they get better with time, no matter the names.

The models are already available via API and in OpenAI’s playground, and won’t be available in the user-friendly ChatGPT UI—at least not yet.

Edited by James Rubin