I’ve noticed a significant drop in Astra’s quality.
At launch, Astra delivered truly precise iterations without needing retries. Today, I have to say it’s on par with SOL.
Same task, same level of understanding—simple commands, yet the task isn’t completed. It requires repeated attempts and explanations.
So, has the quality dropped?
I noticed the same thing today. I also ran into this last Sunday - Astra suddenly felt noticeably worse, but after some time the intelligence seemed to return to normal.
Right now, the current Astra model feel much closer to Sol to me.
I’ve noticed something similar with some AI tools, although it’s hard to tell whether it’s an actual quality drop or changes in the model, routing, system behavior, or task handling.
What stands out to me is the need to repeat simple instructions that previously worked on the first attempt. For these kinds of tools, consistency is almost as important as raw capability.
It would be interesting to compare the exact same prompts and tasks from launch vs. now. That would give a better idea of whether the quality has genuinely changed or if it’s more task-specific.
I’m in a very similar boat with Astra recently. I was having it walk me through something that I was trying to learn how to implement, and it could not even get a basic uv command for installing a package right, and then botched commands associated with that package after reading its documentation. I’ve also noticed a decrease in speed along with the decrease in efficacy.
Same prompt, same task, same GPT-6 Astra model on High, same setup. Five days ago, Astra produced a genuinely good SVG - properly recreated as vector content, not just a PNG from the reference embedded inside an SVG.
Today, with everything kept the same, the result was dramatically worse. The difference is easy to see without any subjective benchmarking.
What makes it even more suspicious is the runtime: five days ago Astra spent around 2m 52s on the task. Today it returned a result in only 36s. I also ran the same task with Sol, and even Sol produced a better SVG than the current Astra result.
So at least in this case, it doesn’t look task-specific. Same model, same thinking effort, same prompt, same environment - completely different level of capability.
5 days ago:
Today:
