OpenAI unveiled GPT-4.1 on Monday, a trio of latest AI fashions with context home windows of as much as a million tokens—sufficient to course of total codebases or small novels in a single go. The lineup consists of normal GPT-4.1, Mini, and Nano variants, all concentrating on builders.
The corporate's newest providing comes simply weeks after releasing GPT-4.5, making a timeline that makes about as a lot sense as the discharge order of the Star Wars films. "The choice to call these 4.1 was intentional. I imply, it's not simply that we're dangerous at naming," OpenAI product lead Kevin Weil stated through the announcement—however we’re nonetheless looking for out what these intentions had been.
GPT-4.1 reveals fairly fascinating capabilities. In response to OpenAI, it achieved 55% accuracy on the SWEBench coding benchmark (up from GPT-4o's 33%) whereas costing 26% much less. The brand new Nano variant, billed as the corporate’s "smallest, quickest, most cost-effective mannequin ever," runs at simply 12 cents per million tokens.
Additionally, OpenAI gained't upcharge for processing huge paperwork and really utilizing the a million token context. "There isn’t a pricing bump for lengthy context," Kevin emphasised.
The brand new fashions present spectacular efficiency enhancements. In a stay demonstration, GPT-4.1 generated an entire net software that might analyze a 450,000-token NASA server log file from 1995. openAI claims the mannequin passes this check with almost 100% accuracy even at million tokens of context.
Michelle, OpenAI's post-training analysis lead, additionally showcased the fashions' enhanced instruction-following skills. "The mannequin follows all of your directions to the tea," she stated, as GPT-4.1 dutifully adhered to advanced formatting necessities with out the standard AI tendency to "creatively interpret" instructions.
How To not Rely: OpenAI's Information to Naming Fashions
The discharge of GPT-4.1 after GPT-4.5 looks like watching somebody depend "5, 6, 4, 7" with a straight face. It's the most recent chapter in OpenAI's weird versioning saga.
After releasing GPT-4 it upgraded the mannequin with multimodal capabilities. The corporate determined to name that new mannequin GPT-4o ("o" for "omni"), a reputation that might be even be learn as “4 zero” relying on the font you utilize
Then, OpenAI launched a reasoning-focused mannequin that was simply referred to as “o.” However don’t confuse OpenAI’s GPT-4o with OpenAI’s o as a result of they don’t seem to be the identical. No one is aware of why they picked this identify, however as a basic rule of thumb, GPT-4o was a “regular” LLM whereas OpenAI o1 was a reasoning mannequin.
A couple of months after the discharge of OpenAI o1, got here OpenAI o3.
However what about o2?—Nicely, that mannequin by no means existed.
“You’d suppose logically (our new mannequin) perhaps ought to have been referred to as o2, however out of respect to our pals at Telefonica—and within the grand custom of open AI being actually really dangerous at names—it's going to be referred to as o3,” Sam Altman stated through the mannequin’s announcement.
The lineup additional fragments with variants like the traditional o3 and a smaller extra environment friendly model referred to as o3 mini. Nonetheless, additionally they launched a mannequin named “OpenAI o3 mini-high” which places two absolute antonyms subsequent to one another as a result of AI can do miraculous issues.In essence, OpenAI o3 mini-high is a extra highly effective model than o3 mini, however not as highly effective as OpenAI o3—which is referenced in a single chart by Openai as “o3 (Medium),” appropriately. Proper now ChatGPT customers can choose both OpenAI o3 mini or OpenAI o3 mini excessive. The conventional model is nowhere to be discovered.
Additionally, we don’t need to confuse you anymore, however OpenAI already introduced plans to launch o4 quickly. However, after all, don’t confuse o4 with 4o as a result of they’re completely not the identical: o4 causes—4o doesn’t.
Now, let’s return to the newly introduced GPT-4.1. The mannequin is so good, it’ll kill GPT-4.5 quickly, making that mannequin the shortest residing LLM within the historical past of ChatGPT. "We're saying that we're going to be deprecating GPT-4.5 within the API," Kevin declared, giving builders a three-month deadline to change. "We actually do want these GPUs again," he added, confirming that even OpenAI can't escape the silicon scarcity that's plaguing the business.
At this charge, we're sure to see GPT-π or GPT-4.√2 earlier than the yr ends—however hey, at the least they get higher with time, regardless of the names.
The fashions are already obtainable through API and in OpenAI’s playground, and gained’t be obtainable within the user-friendly ChatGPT UI—at the least not but.
Edited by James Rubin
