Why GPT-5 is the Best Worst Coding Partner I’ve Ever Had
By - Bharat Sharma 2:42 am 20th August 2025
We all know how quickly the world of AI is changing. It feels analogous to the golden age of Physics, where each new day brought a theory that reshaped our understanding of the universe—new discoveries, revolutionary ideas, and a fair amount of debunking of long-held beliefs since Newton.
Now, the same is happening with AI—particularly with large language models (LLMs), which are a subset of artificial intelligence. Each day brings a new model, a better one, a more efficient one—and sometimes our previous understanding gets overturned, just as it did earlier this year with the release of DeepSeek.
And now we have another leap forward. OpenAI—often called the pioneers of commercial agentic AI—recently announced and rolled out their latest flagship model: GPT-5. After two weeks of daily use, here is my very honest review. I’ll cover what I loved, what frustrated me, and whether I think GPT-5 is truly a qualitative improvement over previous models—especially in the context of my project: building a ray tracer physics engine for computer vision tasks.
First Impressions
Let’s get this out of the way first: GPT-5 is great.
It thinks harder, it explains longer, and its answers feel deeply qualitative when it comes to coding problems and brainstorming next steps. Compared to previous models, I’ve found it sharper, more nuanced, and more “humanlike” in reasoning.
But as with every breakthrough, there are growing pains.
Major Issue: Windows Performance
The biggest problem I’ve faced is not the model itself, but how it runs on Windows.
On the Plus subscription, if a project contains multiple chats and one chat goes beyond a certain length, the app begins to lag severely. And when I say severely, I mean the entire page stops responding. The “thinking longer” indicator shows up, but after a few seconds the page freezes completely. I’ve had to reload or restart the app in a different tab to recover.
Initially, I thought this was a Chrome issue, but the exact same (if not worse) problem happens in the native ChatGPT Windows app. During long coding sessions—where quick back-and-forth troubleshooting is essential—this slowdown was extremely frustrating.
Interestingly, this issue doesn’t exist on my iPhone 16 Plus, where the app runs smoothly. This suggests a Windows compatibility or optimization issue that OpenAI hasn’t fully addressed yet.
Where GPT-5 Shines
Despite these hiccups, GPT-5 shines where it matters most: deep learning and coding support.
I’ve always wanted to explore graphics programming and physics engines, being a lifelong video game fan. When I finally decided to build my own ray tracer from scratch, I didn’t know C++ beyond the basics. I had some Java OOP experience, but that only took me so far.
The real challenge wasn’t just coding—it was the physics. Beyond high-school optics, I didn’t know much about how light really works. That’s where GPT-5 became invaluable.
It helped me translate physical theories into code while also teaching me the math and physics behind them—sometimes in extreme detail, and other times in simplified explanations when things got too complex. The “deep thinking” abilities of GPT-5 made it feel like a patient tutor that could flex between technical rigor and beginner-friendly clarity.
The coding capabilities are also outstanding. Compared to Gemini 2.5 Pro—which I found good but inconsistent—GPT-5 produces cleaner, higher-quality code with fewer prompts. It’s the most productive AI coding assistant I’ve used so far, combining correctness with readability.
A Leap Forward—But Which Kind?
So, is GPT-5 really a leap forward?
Qualitatively, yes.
- More nuanced answers.
- Deeper logical reasoning.
- Cleaner, more efficient code generation.
- Humanlike explanations that adapt to your level.
As a daily ChatGPT user for over two years, I can confidently say this is the best model yet in terms of quality.
Quantitatively, not quite.
On paper, GPT-5 is a monster in terms of raw parameters and test scores. But when it comes to time to task completion, it struggles—because of the optimization and performance issues I described on Windows. It’s not about “thinking longer” per se—it’s about the delays, reloads, and interruptions required just to retrieve answers. That kills productivity.
Verdict
- For quality: GPT-5 is unmatched.
- For productivity: GPT-4o is still my go-to, because it runs smoother, faster, and integrates better into my workflow without lag.
- For developers like me: GPT-5 is incredibly exciting. With some bug fixes and optimizations, it will easily take the crown as the best everyday model.
Personally, I love GPT-5. It has already taught me more about C++ and physics than I ever thought I’d learn in two weeks. But until OpenAI irons out the optimization issues, I’ll keep relying on GPT-4o for uninterrupted daily work—while turning to GPT-5 when I need deep, detailed, high-quality thinking.