After a great deal of anticipation, Sam Altman and the OpenAI team finally released ChatGPT-5 at the OpenAI Summer Update yesterday. While there鈥檚 been a fair bit of subtle hinting that the launch was imminent, the last indication was a little more on the nose, shall we say 鈥� OpenAI published a Tweet stating simply, 鈥淟IVE5TREAM THURSDAY 10AM PT鈥�. Sure, they didn鈥檛 explicitly state that the live stream was all about GPT-5, but I鈥檇 like to think that, while some people may have assumed some silly intern made a big old career-ending typo, the majority of the tech community took careful note of that curious 鈥�5鈥�.
Now, there鈥檚 a lot to be said about the Summer Update 鈥� from all of the exciting new features and upgrades of GPT-5 to the fact that I am now about 85% certain that Sam Altman is an actual robot. They announced big changes in coding capabilities, impressive new applications in health and medtech as well as the promise of far fewer instances of hallucinations 鈥� all in all, some pretty monumental steps forward.
So, for those who had better things to do yesterday than be simultaneously impressed and perplexed 鈥� and also kind of lulled into a bizarre trance by Altman鈥檚 carefully scripted monologues and the uncomfortable pseudo-off-the-cuff dialogue between OpenAI engineers and execs 鈥� we鈥檝e put together a little explanation of what you missed. Because, as much as we may make jokes about the employees鈥� cringeworthy presentation skills 鈥� no shade whatsoever, they鈥檙e clearly insanely skilled and experienced technical experts 鈥� the work they鈥檝e done on GPT-5 is pretty impressive, and it seems like a pretty darn big move towards industry-wide ambitions to achieve GenAI.
听
GPT-5: The Highlights
听
For those who want a quick snapshot of what鈥檚 new and what was announced yesterday, here鈥檚 a quick summary fo the main updates, changes and new features of OpenAI鈥檚 brand new ChatGPT-5.
听
Core Upgrades to the Model
听
These are the primary upgrades you can expect shifting from GPT-4 to GPT-5:
听
- Unified AI System: GPT-5 can easily, quickly and efficiently switch between 鈥渇ast/basic,鈥� 鈥渄eep reasoning鈥� and 鈥渕ini/nano鈥� modes depending on the complexity of your request.
- Sharper reasoning: Significant improvement in multi-step logic, problem solving and context analysis.
- Far Fewer Hallucinations:听GPT-5 will hallucinate less than ever before, and it鈥檚 also less 鈥渄eceptive鈥�.
- State-of-the-Art Performance Across a Variety of Domains: It鈥檚 especially strong in coding, maths, writing, health advice and visual understanding.
- Improved Accuracy: More reliable, factual responses with better source grounding.
听
These are the promises, and they鈥檙e pretty monumental.
听
New UX Features
听
User experience is a significant component of how much people like using AI models, so these improvements and new UX features could go a long way in making GPT-5 a new favourite:
听
-
Personalities: Choose from different conversational 鈥渟tyles鈥� for the AI (e.g., friendly, concise, professional).
-
Customisable interface: New themes and layout options for personalisation.
-
Voice Upgrades: Far more natural-sounding voices, faster response in voice mode and better back-and-forth flow.
-
Tool Integrations: Direct connection to Gmail, Google Calendar and other services for scheduling, email drafting and workflow help.
听
These new UX features sound cool, and to be honest, judging from the demos, they look pretty cool too.
听
Specialised Capabilities
听
This is undoubtedly where most people were focused, and for good reason. These changes are pretty big, and OpenAI鈥檚 focus 鈥� especially their kind of surprising emphasis on health-related issues 鈥� was fascinating.
听
- Enhanced Coding: It鈥檚 cleaner, there鈥檚 more modular code generation, better debugging, improved adaptation between languages and stronger context memory for multi-file projects. It鈥檚 impressive, there鈥檚 no way around it.
- Advanced Healthcare Guidance: Smarter, more context-aware health information but it came with very, very clear disclaimers that this is听not a replacement for medical professionals.
- Multimodal Improvements:听Better image understanding and integration with text-based reasoning.
These three points are pretty important, and the focus is intriguing 鈥� coding is no surprise, but health (especially their interview with users who personal use of GPT-5 during a particularly difficult personal medical situation) was not on my bingo card for yesterday鈥檚 presentation.
听
Access and Availability
听
So, the one thing everyone wants to know 鈥� what about access? How good is the free version going to be and what are the options for upgrades? Well, the unpaid versions of GPT-5 are actually pretty generous, in my opinion, which I鈥檓 sure will be good news to plenty of users.
听
- Free Access for All: GPT-5 available to all ChatGPT users, with usage limits for free accounts, of course.
- Pro Tier Advantages: Pro subscribers get extended usage, priority speed and access to the most advanced 鈥減ro reasoning鈥� mode 鈥� that鈥檚 the big, exciting part!
- Mini/Nano Versions:听Lightweight variants for mobile and low-resource situations, with faster responses.
听
All in all, OpenAI has made access to GPT-5 pretty inclusive 鈥� I mean, anyone can use GPT-5 for free, and even though their are limits on how much you can use it, that鈥檚 pretty useful.
听
AI Safety
听
Unsurprisingly, given the fact that safety and security have always been a massive topic in AI and only seem to be becoming more prominent, the Summer Update did focus on the issue with a comprehensive explanation and a decent demo. Here are the highlights:
听
- Inputs Are Safer and More Aligned: Reduced bias, fewer misleading answers and less 鈥渟ycophancy鈥� (randomly agreeing without thinking).
- GPT-5 Framed As a 鈥淧roactive Thought Partner鈥�: Clear nods towards moving closer toward artificial general intelligence (AGI) but not AGI yet.
- Refined Memory:听 GPT-5 keeps relevant context without bloating responses or repeating information.
听
All important, with the obvious emphasis on the specific reference to a 鈥減roactive thought partner鈥� and not making any wild claims about AGI.
听
Main Takeaways from OpenAI鈥檚 Summer Update
听
So, this all sounds cool, and futuristic and oh-so-impressive, but what do the experts think about GPT-5? Has it lived up to the hype?听I spoke to Paul Armstrong from the TBD Group, Founder and creator of 鈥榃hat Did OpenAI Do This Week?鈥�, to find out what he has to say about yesterday鈥檚 updates, and all in all, his response was equally optimistic and measured. In summary, Armstrong definitely thinks that GPT-5 is a step up, but, as always, he鈥檚 cautious about some potentially overreaching statements, and I have to agree.
According to Armstrong, 鈥淕PT鈥�5 is a systems-level upgrade, not just another model release. The real advance isn鈥檛 in language but in orchestration. The router architecture allows it to manage its own workload, deciding when to respond quickly or reason more deeply. That shifts user expectations. You鈥檙e not picking the model anymore. The model is picking its own brain.
Still, claims of 鈥淧hD-level intelligence鈥� are misleading. GPT鈥�5 is better tuned, more honest when it fails, and slightly more cautious. But it is still a static, probabilistic output machine. It doesn鈥檛 think, it predicts. The main difference is that now it knows when to pull back instead of bluffing.鈥�
One of the major questions most people (myself included) have about GPT-5 is simple 鈥� how does it compare to GPT-4? While GPT-4 may have been the shiny new thing not too long ago, there鈥檚 no doubt about the fact that GPT-5 outdoes it on loads of different fronts.
鈥淕PT鈥�4 was erratic and brittle under pressure. GPT鈥�5 is more stable, less prone to collapse when prompted aggressively or asked to perform multi-step logic. It avoids the embarrassing failures that defined GPT鈥�4鈥檚 limitations in professional use. The leap isn鈥檛 in cognitive architecture, it鈥檚 in constraint management. GPT鈥�5 reduces hallucination, trims sycophancy, and defaults to silence when uncertain. That makes it more useful in practice. The underlying model is still static and narrow in scope, but the packaging now suits operational environments where reliability matters more than novelty,鈥� says Paul.
All in all, he thinks, 鈥渦sers are going to love it,鈥� and it鈥檚 hard to contest this one.
Now, we鈥檝e got the gist of general reactions to GPT-5, I鈥檇 like to think. But, what about two of the main bits 鈥� in my eyes, two standout features and points of emphasis from the live stream 鈥� that is, applications in health and coding? I asked Armstrong about both, and here鈥檚 what he had to say.
听
What Does GPT鈥�5鈥檚 Increased Coding Capability Mean for AI?
听
鈥淕PT鈥�5 shifts coding from execution to intent. You鈥檙e no longer writing code, users are simple issuing design directives. The demos were slick and aimed at lots of different levels which was smart. OpenAI showed that it handles more than just syntax and logic, but also stylistic and architectural elements. Now that鈥檚 an inflection point; prompting becomes programming will open doors to a lot more people and side hustles whilst simultaneously changing the role of the developer.
Coding is not unvaluable by any means, but what鈥檚 clear is the direction of travel. These models will just get better and better, and developers will likely learn how to supervise systems, not script them.鈥�
听
What Do You Think of GPT鈥�5 in the Medical and Health Context, Especially Given the Privacy and Confidentiality Concerns?
听
鈥淗ealth was the interesting push in the presentation. Mentioned early and all the way through, with an emotional element you don鈥檛 often see in these events. All that said, deploying GPT鈥�5 in health contexts is premature. The model performs well on benchmarks, but benchmarks aren鈥檛 medicine. The model doesn鈥檛 reason diagnostically, and it is imitating authority. That鈥檚 not competence, that鈥檚 pattern-matching with confidence and will be problematic.
This said, there鈥檚 a lot to like with how it can empower people to have better exchanges with doctors before, during and after appointments. Still, the dangers are clear: until GPT鈥�5 can be independently audited, verified, and governed in clinical contexts, the use cases should be considered helpful at best. However, again, you can see the direction of travel, the model will only get smarter and more attuned to answer questions without adding ambiguity or changing the doctor/patient trust relationship.听 鈥�
听
That鈥檚 a Wrap on GPT-5 (for Now)
听
Well, there you have it 鈥� OpenAI鈥檚 highly-anticipated ChatGPT-5 model is now out in the world (well, it鈥檒l be totally out there by the end of next week), ready to be tested and enjoyed in a plethora of new and exciting applications.
As always, there are a few things that may be a little overhyped and a handful of statements that give rise to a 鈥渨ell, maybe not quite鈥� kind of response, but all in all, I鈥檇 say GPT-5 looks pretty darn good and it鈥檚 going to interesting to see not only what people do with it, but where OpenAI plans to go after this.
So, get testing (it鈥檚 free!) and let us know what you think about it.
