The discussion around Astra and OpenAI’s next generation of models is exciting. I understand why people want more capable models, and I am not against OpenAI continuing to develop them.
However, I think there needs to be a better balance between launching new models and improving the ChatGPT experience people already use every day.
There are still bugs, inconsistent behaviors, and features that do not always work well together. Some of these problems have affected my own use for months, even after I reported them.
Instruction-following is still unreliable
One of my biggest frustrations is that ChatGPT does not always follow instructions properly, especially inside Projects and custom GPTs.
Even when the instructions are clear, the model may ignore part of them, follow them only sometimes, or stop following them later in the conversation. This makes it difficult to create reliable, repeatable workflows.
Projects and custom GPTs are supposed to provide more structured and consistent experiences. But when their instructions are not followed consistently, much of that value is lost.
Features can feel disconnected
The experience can also vary depending on the platform, device, mode, account, or subscription.
Features such as voice mode, screen sharing, and Projects do not work together. Some features may only be available in certain modes or on certain devices, while others behave differently across web, mobile, and desktop.
I understand that some of these differences may be intentional limitations rather than bugs. Still, from a user’s perspective, ChatGPT can sometimes feel like several separate products instead of one connected experience.
Updates can make things less predictable
Another frustrating issue is that behavior can change after updates.
Something that worked well before may suddenly become less reliable. Instruction-following, context handling, memory, or response quality can vary between models and even between sessions.
I understand that an AI system will never produce exactly the same result every time. However, users still need a reasonable level of consistency, especially when they rely on ChatGPT for repeated or structured tasks.
Stability may be more valuable than another new feature
For users who rely on ChatGPT regularly, one of the most useful improvements would be making the existing system more reliable.
That could include:
- better instruction-following in Projects and custom GPTs;
- fewer regressions after updates;
- more consistent behavior across platforms and devices;
- clearer information about known problems and limitations;
- faster fixes for issues that users repeatedly report;
- better integration between existing features.
These improvements may not sound as exciting as announcing a new model, but they could make a much bigger difference in everyday use.
Reliability is part of intelligence
A model can be extremely capable, but if it does not follow instructions or behaves unpredictably, it becomes difficult to trust.
In real-world use, reliability matters just as much as raw capability. A slightly less advanced model that consistently follows instructions may be more useful than a more powerful model that frequently ignores them.
Reliability should therefore be treated as a major part of model quality, not as a separate issue.