After dozens of teasers, OpenAI finally launched it today GPT-5。
Not long after launch, it refreshed its score on the lmarena.ai leaderboard and shot straight to first place:
- Solidly first in text, Web development, and visual arenas;
- It also topped the charts in hard prompts, coding, math, creativity, long queries, and more.
But awkwardly, the chart generated by GPT-5 at the launch event seemed to flop. 52.8 is greater than 69.1? 30.8 and 69.1 are the same height?
Is this what Sam Altman calls OpenAI's smartest model ever?
I tested it right away, and aside from coding, it's pretty weak... I just want to say: give me back4o。
01. Firsthand Testing
case 1 A Classic Tricky Question
First, ask a classic math question:
Prompt: Which is larger, 9.9 or 9.11?
GPT-5 answered: 9.11 is larger than 9.9, because when comparing numbers, you first compare the integer part (both are 9), then compare the decimal part, and 11 > 9.
If you're asking about dates, then September 11 is later than September 9.
case 2 Programming
Prompt: Write a JavaScript code that shows a small ball bouncing inside a rotating regular hexagon. The ball should be affected by gravity and friction, and it must bounce realistically off the rotating walls. Implement it with JavaScript and HTML.
The overall effect is okay; you can tell it's affected by gravity and friction, but the bouncing effect isn't particularly prominent.
case 3 Image Generation
Prompt: Translate the content in the image into Chinese. It needs to be easy to understand and engaging, without changing the original meaning, and do not add content out of thin air.
Let's compare; which translation do you all think is better?
Gemini 2.5 Pro：
Doubao:
GPT‑5：
In my view, Gemini 2.5 Pro translated it best; it's very fluent and very human. Doubao is next; although there are no major errors in the translation, it lacks a bit of human touch. GPT 5 is about the same as Doubao, but compared with Gemini 2.5 Pro, the gap is quite obvious.
case 4 Frontend
Prompt: Generate aPomodoro timerweb app with the following features: - Goal: timer + session records. - Features: custom duration, auto-start option, session chart (SVG), sound alert (beep), dark mode. - The interface should be suitable for daily use and give users some energy!
The Pomodoro timer webpage generated by GPT‑5 has better overall aesthetics, and the functions of buttons such as start and pause are complete, but this big ring on the page... I don't understand it.
The page generated by Gemini 2.5 Pro has very complete functionality, but it's rather crude; the layout and color scheme aren't as good-looking as GPT‑5, but you can't deny that it works!
02. Some Sharing
Overall, after testing, I don't know whether GPT‑5's hallucinations have decreased, but its intelligence really seems to have decreased. Not only has its copywriting level not improved, it can't even compare to the previous GPT‑4o.
Its programming ability is also just average. Maybe it's because I didn't use very complex prompts to test it...
The frontend aesthetics feel pretty good, but there are still some small bugs.
These were all One Shot, without multiple rounds of testing, but it can't be that I hit a landmine every single time, right?
Finally, I just want to say, this launch event was a waste of time. GPT, give me back my sleep!
GPT‑5 is now available on the official website for everyone to use. The free version allows 10 messages every 5 hours, and the Plus version allows 80 messages every 3 hours.
If you don't believe it, you can go try it yourself.
Original link:Is 9.11 greater than 9.9? GPT-5 gets even math problems wrong—did it really get an upgrade?