Altman and OpenAI have discovered how important personality is for chatbots. (Picture: generated)In an update on the GPT-5 rollout today, CEO Sam Altman launched a bevy of new features. Like a longer context window for GPT-5-thinking, and changes to the model picker.
The launch of GPT-5 has been a little bumpy, at best. The first reaction was from users missing GPT-4o, which was quickly returned — but what about the other «legacy models?»
Almost all models returned
They are all coming back, and as per now the model picker lists GPTs 4.1, o3 and o4-mini for paid users. The only one missing from before the GPT-5 launch is GPT-4.5, which Altman says «costs a lot of GPUs.»
X.ai feels left out in the App Store, and claims it is because of Apple’s special relationship with OpenAI. (Picture: X.ai)After noticing the Grok app was the 5th choice on the App Store’s top list and ChatGPT was number one, Musk reckons they must be playing their lists to OpenAI’s benefit.
He now threatens an antitrust lawsuit, claiming that ChatGPT is on «literally every list where you [they] have editorial control,» like their «Must-Have Apps» section.
Only «a small percentage» get delusional from ChatGPT use, Altman says. (Picture: Adobe)As more and more publications are digging into people getting delusional from AI use, being led down rabbit holes or thinking they are superhuman, the CEO of OpenAI addressed the topic today.
In a lengthy x.com post, Altman considers the issues as «edge cases,» but welcomed both attachment and using ChatGPT as a kind of «life coach.»
Recently, OpenAI announced a wellness update to reduce sycophancy and push back against delusions, and the hope is that this can reduce some of the risks:
After an uproar on Reddit, ChatGPT 4o is back by popular demand. (Picture: Generated)UPDATED with how-to: GPT-5 might well function as a PHD-level assistant, but it is less chatty and not as filled with self-affirming responses as 4o, as CEO Sam Altman discovered in his Reddit AMA (Ask Me Anything).
He also said on X.com that «We for sure underestimated how much some of the things that people like in GPT-4o matter to them, even if GPT-5 performs better in most ways.»
Anthropic’s Opus 4.1 is very close to the state of the art, and many users are claiming it’s way better than 4.0. (Picture: Anthropic)Anthropic announces Claude Opus 4.1
In an incremental update that got lost in this week’s headlines, Opus has been «improved across most capabilities» relative to the 4.0 version. It now scores 74.5% on SWE-bench Verified, almost as good as GPT-5. Windsurf says the performance gains are similar to going from Sonnet 3.7 to 4. It’s available now and costs the same as Opus 4.0. Users are also noting a significant improvement.
Google says people are still clicking
After a Pew Research report said users are less likely to click on from AI Overviews in Google, the entire publisher scene erupted and saw doom and gloom on the horizon. They were already seeing fewer clicks from Google in their logs. Now, Google is trying to counter with a happy blog post claiming average click quality has actually increased, and that they are in fact sending more «quality clicks» to publishers than before. Not stats, studies or other underpinning for that, though.
GPT-5 made this game on the first try, and it works pretty nicely, too! (Picture: OpenAI)The latest ChatGPT model excels at coding, writing and health queries — and ends the alphabet soup of models.
—Our smartest, fastest, most useful model yet, with built-in thinking that puts expert-level intelligence in everyone’s hands, says OpenAI on their launch post.
All users, including the Free tiers should be getting new models directly after launch, while Enterprise and Edu users will have to wait a week.
No one believes Thursdays livestream will be anything but this. (Photo: generated)Closing out what insiders have been cooing about as their «Big Week,» the livestream announced for Thursday 10:00 PST/19:00 CET is as sure a bet as you can get — it will be the launch of GPT-5. The announcement is not subtle about it, and even calls it a “live5tream.”
The rumors have hit a high pitch since Altman teased the model in a podcast two weeks ago, and various aliases for advanced models started showing up at some benchmarking sites.
It will also run longer than usual, as there is a lot to cover, says @sama:
ChatGPT usage could shed some serious time by automating routine work. (Picture: Generated, Mark Hillary, CC BY 2.0)OpenAI has partnered with the U.S. General Services Administration (GSA) to provide ChatGPT Enterprise basically for free.
Potentially all of government civilian employees are covered by the deal, reaching some 2—3 million in the current federal workforce. That would be a boon for OpenAI, but more in the sense of prestige than in usage, as ChatGPT currently has around 700 million weekly users.
The agreement could likely cement a full year of entrenched habits, history and preferred use if successful — basically locking in OpenAI as the preferred AI provider across federal agencies.
In Genie 3, the world’s your oyster — and you walk around in it, or mash things up. (Picture: Google DeepMind)The model creates deeply realistic 3D worlds that you can interact with to «achieve goals,» improve education or just have a fun old time.
It builds upon its own previous version, which could only support video in 10-20 seconds at 360p, and on Veo 3 — which makes photorealistic non-interactive videos.
The result is a model that can make 720p video at a smooth 24 fps, understands physics, remembers about a minute back of previous renders for consistency, and creates a 3D world that you can move around in «for a few minutes.»
Two agentic, reasoning models will fit on high en consumer hardware — if you have the specs. (Picture: Adobe)The two models will run on a high-end laptop or a phone, and perform at the level of o4-mini.
Sam Altman says «We believe far more good than bad will come from it,» choosing to release the models after a series of delays and worry about the weights.
After «billions of dollars of research» and extensive red-team testing, they were found no more dangerous than the o3-model and won’t move the needle on chemistry or biology.
Numbers keep rising for ChatGPT, but the mental health updates might be more important. (Picture: Adobe)The milestone comes just four months after they announced 500 million in March, and is four times the volume they had last year.
That number spans all of ChatGPT’s accounts, ranging from free to Pro and everything in between, and is rapidly accelerating, writes CNBC.
Reddit is going all in on AI search for its huge content of 20 billion posts and comments. (Picture: Reddit)If there’s anything Reddit is famous for outside of daily memes, easy news and activism — it is the very fact it is all pedal-driven. As in it is full of human connections, giving tonnes of answers to questions ranging from the trivial to the weird.
The company grew its revenue by 78% year-over-year this quarter, and now has a plan to drive even more traffic to its services; AI.
«We will make the investment to do it,» Cook said. (Picture: Adobe)Add cloud computing and apps, and you have the mindset of CEO Tim Cook at an all hands meeting in Cupertino yesterday, as reported by Bloomberg.
— Apple must do this. Apple will do this. This is sort of ours to grab, the Apple chief said.
He also said during the Q2 earnings call on Thursday that Apple would «invest significantly» in AI projects next quarter.
— We’ve rarely been first, he added, telling employees that — There was a PC before the Mac, there was a smartphone before the iPhone, there were many tablets before the iPad, there was an MP3 player before iPod.
Together, these statements might signal a new era for Apple — attacking the AI market with some sense of renewed urgency, and investing aggressively in the process.
Google’s new model thrashes the benchmarks, gets IMO ranked and excels at creativity. (Picture: Screenshot, Google)This a version similar to the one that won gold in the International Mathematical Olympiad, this time performing at bronze level — a feat no other released model can manage.
Gemini Pro 2.5 Deep Think features «parallel thinking,» letting it work through tens, if not hundreds, of different solutions to a problem simultainiously, test them, revise them or even combine them to return the best answer.
OpenAI has since removed the checkbox that would share your chat with search engines.(Picture: Screenshot, OpenAI)If you clicked this box, your ChatGPT session would be on Google
ChatGPT was just two clicks away from spilling your secrets to Google, an investigation found yesterday. The «Make this chat discoverable» button on the share feature would register the whole chat on search engines. Plenty of people made that mistake, sharing «deeply personal details, including struggles with addiction, experiences of physical abuse, or serious mental health issues.» OpenAI removed the feature shortly after, saying it «introduced too many opportunities for folks to accidentally share things they didn’t intend to.» They are now scrambling to remove «indexed content from the relevant search engines.»
Apple open to mergers and acquisitions in AI space — We’re very open to M&A that accelerates our roadmap, Apple’s Tim Cook said on yesterday’s earnings call. — We are not stuck on a certain size company, although the ones that we have acquired thus far this year are small in nature. He also said Apple was going to «significantly» grow it’s AI investments, after the company reported a 10% increase in revenue — the largest jump since 2021. They are also «making good progress» on personalizing Siri, he said.
Zuckerberg throws shade on open source projects
Mark Zuckerberg of Meta recently posted a manifesto of sorts, mapping out what he sees as a benevolent «personal superintelligence» «for everyone.» In it, he quietly states that superintelligence will pose new safety concerns, and «We’ll need to be rigorous about mitigating these risks and careful about what we choose to open source.» He repeated this in his later earnings call, saying «we kind of wrestle with whether it’s productive or helpful to share that.» See also his original post.
Developers are slightly souring on AI coding A new survey from Stack Overflow shows a significant drop in developer trust in AI coding, with trust in its accuracy dropping from 40% in previous years to 29% in 2025. «Favorability» has also dropped from 72% to 60% on a yearly basis. 52% of developers say they use AI agents in their work, while 72% reject «vibe coding.» The survey was taken with 49,000 worldwide developers. Stack Overflow is no stranger to AI effects, having dropped sharply after the AI coding boom.
