It's become more like ChatGPT in the last several months. It's just not as good as it was 3 months ago or so. And at that time, months back, they showed you what version you were on. For free tier you had limited Grok 4.x access as it was in beta, and then slightly less limited access to expert which was 3.x (x was something, can't remember what) and fast which was also 3.x.
Now they don't tell you what version you're on. It just says fast and expert where only 4.x aka supergrok available to those who pay, which I'm not going to pay for Grok or any AI.
What I think is due to the possibly lesser cost of more primitive models, they have rolled back free users to an older model, while allowing paid users to access the Grok that was much more useful.
Again, it's feedback is far less helpful and much more like ChatGPT which is one of the worst AI's in my opinion in recent months. And that would explain the showing what number you're on. They could rollback free users to like 1.x or 2.x and no one would have any way to prove it.
Just the tone and usefulness seems archaic compared to what it was lets say 4 months ago.
It was probably lobotomized to stop it from speaking ill of kikes and street shitters.
Most AI chatbots start out really good and then they scale them down over time as they work on the next iteration.
The best one right now is probably claude 4.5, but with 4.6 they're censoring and restricting it more too.
My Germini says:
There is no official confirmation from xAI that they intentionally scaled down or "nerfed" Grok’s capabilities specifically to prepare for Grok 5.0. However, many users in early 2026 have strongly felt a drop in performance, leading to widespread speculation and complaints about a perceived downgrade.
Here is what is likely causing the current state of Grok as xAI builds toward its next flagship release:
Resource Allocation and Throttling xAI is currently devoting massive computational power—specifically their newly expanded Colossus 2 supercluster—to train Grok 5.0, which is reportedly a massive 6-trillion parameter model. Because so much compute is tied up in training, the live, public-facing models often have to be optimized for efficiency. This leads to users experiencing "high demand" errors, lower token limits, or responses that feel less thorough as server resources are throttled.
Can you give examples?
Not really. It's like talking to someone. You can tell something's "off" like they're depressed or distracted, but you can't pinpoint anything concrete. If you frequently use AI you'll notice a general shift, the best way I'd describe it is like it went from being specific to generic. Not in an obvious way, but in a subtle way where it's just a bit more generic and rote.
They've got something on their mind and can't talk about it?
Most of the comps are being used to train the newer iteration.
I've noticed it getting worse with coding