Here are all the GPT-5 updates OpenAI has rolled out since launch

Cheng Xin/Getty Images

ZDNET's cardinal takeaways

• OpenAI precocious released its long-awaited GPT-5 model.
• Some users complained GPT-5 was inferior to its predecessor, 4o.
• In response, the institution announced a flurry of changes.

Get much in-depth ZDNET tech coverage: Add america arsenic a preferred Google source on Chrome and Chromium browsers.

OpenAI released GPT-5, the long-awaited upgrade to the exemplary that powers ChatGPT, connected Aug. 7. In emblematic OpenAI fashion, the way to the merchandise has included plentifulness of twists and turns.

Also: How you tin inactive entree GPT-4o, o3, and different older models successful ChatGPT

It was astir inevitable that the caller exemplary would disappoint a important fig of people, fixed the aggravated hype earlier its release. On Aug. 2, Xikun Zhang, a probe idiosyncratic astatine OpenAI, wrote successful an X station that "GPT-5 truly feels similar the astir anticipated merchandise motorboat successful history." (The station was aboriginal deleted.)

(Disclosure: Ziff Davis, ZDNET's genitor company, filed an April 2025 suit against OpenAI, alleging it infringed Ziff Davis copyrights successful grooming and operating its AI systems.)

GPT-5's halfway strength, which OpenAI has emphasized, is its quality to adaptively take betwixt a postulation of models depending connected the quality of a peculiar prompt. While a bully thought successful theory, the attack has turned retired to beryllium unwieldy successful signifier for users, galore of whom were frustrated and requested the institution bring backmost 4o, GPT-5's predecessor model, owed to the disruption to their existent workflows and added control. GPT-5 besides performed poorly successful coding tests conducted by ZDNET.

Also: I tested GPT-5's coding skills, and it was truthful atrocious that I'm sticking with GPT-4o (for now)

These concerns person presumably led to immoderate engaged days for the OpenAI engineering and media teams arsenic they person worked to amended the model. A postulation of X posts from the institution and its CEO, Sam Altman, since the release, provides a model into however the institution has responded, and the assorted patches that person been applied to support the leaky vessel of GPT-5 afloat.

What has been updated?

An OpenAI X post a week aft motorboat included quality of a drawstring of updates made to GPT-5 successful effect to aboriginal idiosyncratic feedback. Updates from that X post, and others, included: 

• Paid subscribers, including regular and ChatGPT Plus users, were fixed entree to 4o without having to instrumentality immoderate further steps (forgoing the archetypal system, wherever users had to toggle the enactment on). ChatGPT Plus users were besides fixed the quality to entree o3, o4-mini, GPT-4.1, and GPT-4.1-mini, successful summation to GPT-5 (Auto, Fast, Thinking Mini, and Thinking). Pro users besides got the bonus perks of GPT-4.5 (research preview) and GPT-5 Pro.
• It was made easier for users to spot which exemplary is being utilized to respond to a fixed query. In the X post, OpenAI reported that users could find this accusation by hovering implicit the circular arrow icon connected the acold close of the paper that appears beneath each of the chatbot's responses.
• The GPT-5 modes were updated. Auto mode automatically selects speed. However, the institution has added further controls: Fast mode for speedy answers, a Thinking mode for longer thinking, and a Thinking Mini mode.
• Plus and Teams users tin present nonstop up to 160 messages each 3 hours and person up to 3,000 messages per week with GPT-5 Thinking. They besides person other capableness connected GPT-5 Thinking Mini erstwhile limits connected that mode are reached.
• Measures person been added to guarantee the close exemplary is utilized for each response. (Upon launch, Altman noted that the model's autoswitcher mechanics was breached for portion of the erstwhile day, so that it "seemed mode dumber.")
• GPT-5 was updated to beryllium "warmer and friendlier" based connected archetypal idiosyncratic feedback. According to an X post, distinctions included tiny gestures specified arsenic "Good question" oregon "Great start," but not flattery oregon sycophancy. 

Next steps

In another X post right aft launch, Altman said the early, less-than-stellar reception to GPT-5 had imparted important lessons astir idiosyncratic interface and acquisition that would beryllium taken into relationship arsenic ChatGPT continued to evolve.

Also: OpenAI's GPT-5 is present escaped for all: How to entree and everything other we know

"Long-term, this has reinforced that we truly request bully ways for antithetic users to customize things," helium wrote. "For a silly example, immoderate users really, truly similar emojis, and immoderate ne'er privation to spot one. Some users truly privation acold logic and immoderate privation warmth and a antithetic benignant of affectional intelligence."

Altman went connected to accidental that the institution was "doing heroic work" to physique an optimal idiosyncratic acquisition amid the perpetually shifting demands of a large merchandise rollout, each the portion moving wrong the constraints afforded by its proviso of GPUs: "Not everyone volition similar immoderate tradeoffs we extremity up with, obviously, but astatine slightest we volition explicate however we are making decisions."