ZDNET’s key takeaways
- OpenAI launched its long-awaited GPT-5 on Thursday.
- Some customers complained GPT-5 was inferior to its predecessor, 4o.
- In response, the corporate introduced a flurry of adjustments.
OpenAI launched GPT-5, the long-awaited improve to the mannequin that powers ChatGPT, on Thursday. In typical OpenAI trend, the discharge has included loads of twists, turns, and drama.
It was nearly inevitable that the brand new mannequin would disappoint a big variety of folks, given how intensely it was hyped main as much as its launch. On Aug. 2, Xikun Zhang, a analysis scientist at OpenAI, wrote in an X publish that “GPT-5 actually appears like essentially the most anticipated product launch in historical past.” (The publish was later deleted.)
(Disclosure: Ziff Davis, ZDNET’s guardian firm, filed an April 2025 lawsuit towards OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI programs.)
Positive sufficient, GPT-5 has debuted extra with a whimper than a bang.
GPT-5’s core energy, which OpenAI has gone to nice lengths to emphasise, is its skill to adaptively select between a group of fashions relying on the character of a selected immediate. Whereas a good suggestion in principle, it is turned out to be unwieldy in follow for customers, a lot of whom are actually requesting out of frustration that the corporate convey again 4o, GPT-5’s predecessor mannequin. GPT-5 additionally carried out poorly in coding exams performed by ZDNET.
It did not assist that a number of charts displayed inaccurate info throughout a livestream presentation of GPT-5 on Thursday, a slip-up that the web has taken to deriding because the “chart crime.” One bar graph evaluating the brand new mannequin’s efficiency on the SWE-bench check, for instance — which gauges fashions’ coding abilities — confirmed a relatively low rating for GPT-5 being displayed inside a disproportionately increased bar.
All of this has presumably led to a busy handful of days for the OpenAI engineering and media groups as they cope with the fallout of this newest snafu. A set of X posts each from the corporate and its CEO, Sam Altman, since Thursday, supplies a window into how the corporate has responded, and the varied patches which were utilized to maintain the leaky ship of GPT-5 afloat.
What has been up to date?
A Friday X publish from Altman included a string of updates that had been made to GPT-5 in response to early person suggestions. These included:
-
Doubling the speed restrict for GPT-5 for ChatGPT Plus subscribers. In keeping with OpenAI’s web site, Plus customers can now ship as much as 160 messages each three hours through GPT-5. As soon as that restrict has been hit, the mannequin will mechanically swap to a extra restricted mini mannequin till the utilization time window resets. OpenAI additionally famous that this was “a brief enhance and can revert to the earlier restrict within the close to future.”
-
Permitting some paid subscribers to proceed to entry and use 4o. In a X publish on Friday, OpenAI wrote that ChatGPT Plus and Workforce customers can entry the mannequin via an choice studying “present legacy fashions” in settings. The identical publish added that GPT-5 had been efficiently rolled out to all ChatGPT Plus, Professional, Workforce, and free customers.
-
Making it simpler for customers to see which mannequin is getting used to reply to a given question. In the identical Friday X publish, OpenAI stated customers can now discover this info by hovering over the round arrow icon on the far proper of the menu showing beneath every of the chatbot’s responses.
-
Including measures to make sure the fitting mannequin is used for every response. (Altman famous in his Friday publish that the mannequin’s autoswitcher mechanism was damaged for a part of the day past, in order that it “appeared method dumber.”)
Subsequent steps
In one other X publish on Friday, Altman stated the early, less-than-stellar reception to GPT-5 had imparted essential classes round person interface and expertise that might be taken into consideration as ChatGPT continued to evolve.
“Lengthy-term, this has bolstered that we actually want good methods for various customers to customise issues,” he wrote. “For a foolish instance, some customers actually, actually like emojis, and a few by no means wish to see one. Some customers actually need chilly logic and a few need heat and a special sort of emotional intelligence.”
Altman went on to say that the corporate was “doing heroic work” to construct an optimum person expertise amid the consistently shifting calls for of a significant product rollout, all of the whereas working throughout the constraints afforded by its provide of GPUs: “Not everybody will like no matter tradeoffs we find yourself with, clearly, however no less than we’ll clarify how we’re making selections.”
