I switched back to Sol.

After using GPT 6 Astra for a few days, I switched back to 5.6 Sol. Not because Astra isn't great, it is amazing, at many things I have tried. The reason I switched back is I still want my tokens to last through the week.

I am not building Rick Astley's animation in Blender or asking it to paint a portrait in MSPaint. A large portion of my workload is simple but tedious, sometimes challenging.

The thing is with Astra, there's really no difference between fixing some small things and building a really difficult feature. I am already using the lowest reasoning level, but even with that, after two days I'm already halfway down. At this rate, I won't be able to do any work after four days, which never happened to me when I was using Sol.

Now I did switch back, and after using Astra for a long stretch of time. I do find the execution speed is drastically slower, but for many tasks I'm running. I'm not really feeling any difference in quality between Sol and Astra.

Neither model can produce code without any problems.

I'm working on an established codebase where creativity doesn't really do me good. Reliability and precision matter most.

And because of that, I find using the older model fit my needs much better.

There is another hidden benefit of using the older Sol model, which is the slowest speed actually works in my favour, because I can actually multitask and context switch between a few different threads. The new astra model works quite fast and doesn't really allow me to switch between a few active threads without fumbling around to catch up with the model.

So for me right now, if I'm working on a problem that's super difficult to solve, or I need absolute best quality output(for instance some really core part of the code base). Or I'm in a hurry I need to fix something serious really quickly. I might choose Astra for those tasks.

Otherwise, I'm pretty content with the older model which gives me time to think, time to check the result and allowing me to juggle between different threads without worrying about wasting time.

But on the other hand, if money is not object, if I have unlimited tokens, of course I will use Astra for everything. It's faster, smarter, and more responsive, but right now most of the time is an overkill even with the lowest reasoning.

Now, a friend said to me the only problem with using less capable models, which I agree, is when something breaks, you never know if it is because of the lower models capability, or the problem is really difficult to solve.

But let's be real. We are not really talking about the difference between Sol and Luna, which is a real, vastly different pair. Sol itself was already so capable, I am not forgetting how powerful it was just because a better one is out.

Anyway, that's my hands on experience with the two models, hope it helps you deciding which one you should use.