AI is literally everywhere these days, and the developers of Anthropic and Gemini and ChatGPT are always clearly working on the next big thing. It’s probably not going to be energy efficient, but it will probably be more sizeable than before.
Case in point, OpenAI’s ChatGPT has a new version that largely promises to be exactly what, as ChatGPT-6 becomes a thing.
Arriving to largely kick off September (and spring, if you celebrate that), ChatGPT’s latest release is so big, the company is calling it close to AGI, or what is known as “Artificial General Intelligence”, an idea that essentially implies an AI can think like a human and match our ability to process information and create. It’s largely hypothetical, but AI designers believe it’s possible, and with ChatGPT’s version six, also known as “Astra”, the company says it’s basically here.
A much bigger platform and one that leaps over what rolled out in GPT-5, it’s a model that is essentially multi-modal in that it can do a lot more. Create documents, build code, convert music to chords and tabs, and attempt to be the be-all end-all we’ve been told AI is supposed to be.
According to OpenAI, it’s faster and better for computer use, creating spreadsheets and presentations, and essentially building out websites and apps to help humans alongside their work. It’s still a tool, though it’s one effectively built to cover more of everything.
And according to tests, it’s up there in capability, with OpenAI noting it beats previous ChatGPT and Claude models, not just delivering faster speeds, but more capability from agents in unfamiliar interactive tasks. That last one is a part of how AGI can be graded, tested in the ARC-AGI-3 test. For Astra, it’s a result that hits 99.9 percent, compared to 7.8 percent with GPT 5.6 and 30.2 percent with Claude Opus 5.
Does that make GPT-6 so smart, it’s basically artificial general intelligence?
At this point, probably not.
Small tests, big results
We’ve only been playing with GPT-6 for a couple of days — basically the extent of time the model has been made available to users — and while there’s little doubting it’s a big deal, automatically smart on the level of matching humans or even able to conceive outside of them, that it doesn’t appear to be.
You can have it turn a brief into something real and usable, or have it convert a song into lyrics and music, or have it generate imagery and art, or analyse information and so on and so on. It works the way you’ve probably seen an AI system before, but it can bring it all together.
But like other AI systems, it can still hallucinate, make arguably off-instruction suggestions, and in general ignore what you’ve asked it to do.
If this is what OpenAI calls artificial general intelligence, it may not have a high expectation for general intelligence on the whole.
A creative conundrum
Part of the problem may extend from the idea that AGI can imitate a human ability to conceive and create. It’s a noble thought, but because AI is trained from a bunch of different things, it is really only blending ideas, as opposed to creating or riffing from it.
While humans can do that as well, and may be inspired by a lot of things, humans can still create without having seen or heard or touched another concept. Ideas can still originate sans-training, something AI can’t yet achieve.
In fact, in our early tests, a prompted push for creative concepts almost always ignored the concepts we were trying to get across, with the system noting it had misunderstood several times. Like other AI systems, you still needed to push it in the direction you wanted it.
When you did that, the results were definitely faster and more capable, but still lacked the human ability to play with the ideas. They’re very technical and incredibly capable at turning a brief into a digital development, driving concept to code. However, human intelligence and capacity for creativity it was not.
Despite that, the system is definitely interesting, and comes with some curious caveats, such as an admission by OpenAI that GPT-6 is more difficult to monitor. That could mean it’s processing more quickly and in more ways, but in an age where AI can say things that are actually wrong or potentially hazardous to mental health, having a system be less able to be monitored isn’t a fantastic outcome.
In terms of what happens next, more pushing, more prodding, and more testing will be the likely steps, and we won’t be alone. It will be curious to see if the model manages to be more energy efficient, but the guess is that’s not going to be the case at all. Rather, this is likely to be just as hefty due to how much it’s doing in the grand scheme of things.
And for people who want to see that level of “how much”, they’ll want a paid GPT account, working on Plus, Pro, Business, or Enterprise level accounts, or even using the API independently. In short, if you see ads on your version, you probably won’t be playing with GPT-6, at least not yet.
