Since the Astra release, ChatGPT Pro has become increasingly unreliable for serious work
I have reached the point where I seriously need to complain about the current state of ChatGPT for heavy professional use.
I use ChatGPT extensively every day for software development, infrastructure work, remote development through RDC, research, automation, and several large ongoing projects. This is not about an occasional bad answer or one broken conversation. I am talking about a very noticeable deterioration in reliability since the Astra release.
Before Astra, GPT-5.6 Sol at xHigh was excellent for my workloads. It was reliable enough that I could hand it large technical tasks and expect it to keep context, work through the problem, use tools, verify its work, and continue until the task was actually finished.
That is no longer my experience.
GPT-5.6 Sol feels significantly worse
Regardless of the reasoning setting, and regardless of whether I use normal Chat or Work, GPT-5.6 Sol now feels substantially less capable and less reliable than it did before the Astra release.
The problems I am seeing constantly include:
- context being lost inside long-running project conversations;
- tasks simply stopping before completion;
- repeated “too many requests” errors;
- tool workflows getting interrupted for no obvious reason;
- conversations falling into loops instead of progressing;
- the model forgetting decisions, files, constraints, or work completed earlier in the same project;
- long technical tasks becoming increasingly unpredictable.
This is not a subtle quality difference.
For my workloads, 5.6 Sol xHigh before Astra and 5.6 Sol xHigh today do not feel like the same level of product reliability.
Astra itself is excellent — but the limits make it practically unusable
The frustrating part is that Astra is actually very good.
Even Astra at Low can be extremely capable. For difficult autonomous work, it often performs better than what I am currently getting from 5.6 Sol.
The problem is that I cannot use it at the volume required for serious daily work.
I am already on Pro 20x, and I can still burn through the available usage in roughly a day when doing real development work.
Since the Astra release, I have already purchased approximately 6,600 additional credits and reset my weekly limit eight times.
That is an absurd amount of additional capacity to have to consume just to maintain the workflow I had before.
So the current situation feels like this:
5.6 Sol has become too unreliable for the work I need it to do, while Astra is good enough to replace it but too heavily constrained to use continuously — even on the 20x plan.
That is an extremely frustrating position for a paying power user.
GPT-6 Pro in normal Chat is also not usable for my coding workflow
Another major problem is GPT-6 Pro when working on coding tasks through remote desktop/RDC.
It repeatedly refuses or interrupts completely legitimate development tasks because of supposed safety restrictions.
What makes this especially frustrating is that I can take the exact same task, switch directly to GPT-5.6 Sol, and the work continues normally.
That makes these interruptions feel less like meaningful safety enforcement and more like false positives that make GPT-6 Pro unreliable for practical software-development workflows.
For serious coding work, I cannot use a model if I do not know whether it will suddenly decide halfway through an ordinary development task that it can no longer continue.
Normal Chat has become unpredictable
The biggest issue overall is reliability.
I no longer know whether a task will:
- finish normally;
- lose context;
- hit “too many requests”;
- stop halfway through;
- enter an endless loop;
- suddenly refuse an ordinary tool action;
- or simply require me to start over.
For casual ChatGPT usage that may be annoying.
For someone using ChatGPT as part of an actual production workflow, it is unacceptable.
I do not mind paying more for compute. I clearly already do.
What I cannot work with is unpredictability.
If I spend money on additional capacity, I need the underlying system to remain stable enough that the work I am paying for can actually be completed.
I am giving this another week
I have been extremely happy with ChatGPT for a long time, which is why this is so disappointing.
Before Astra, GPT-5.6 Sol xHigh was one of the best tools I had ever used for this kind of work.
Right now I feel as though I am being pushed toward Astra because the previous model experience has degraded, while simultaneously being prevented from using Astra at the volume I actually need because of the limits.
I am going to watch this for another week.
If reliability does not improve, I will start moving my projects to Claude and would rather spend $1,000 per month there than continue fighting limits, context loss, interruptions, false safety blocks, and unfinished tasks here.
That is not meant as drama or a threat. It is simply becoming a business decision.
I would much rather stay with ChatGPT.
But I need the product to become reliable again.
What I would like OpenAI to address
I would genuinely appreciate clarification on a few things:
- Has anything changed in GPT-5.6 Sol routing, compute allocation, context handling, or reasoning budgets since the Astra rollout?
- Why have long-running Chat and Work sessions become so much less reliable?
- Why are “too many requests” errors appearing so frequently even for high-tier paid users?
- Are Astra limits expected to remain this restrictive for Pro 20x users?
- Is OpenAI aware of increased context loss and prematurely terminated long-running tasks?
- Why does GPT-6 Pro trigger safety interruptions on ordinary RDC development tasks that GPT-5.6 Sol immediately completes without issue?
- Is there currently an infrastructure, admission-control, or capacity-management problem affecting heavy users?
I do not need unlimited compute.
I need predictable compute, reliable context, and workflows that actually finish.
At the moment, I am paying significantly more while getting a less reliable working environment than I had before Astra.
That is the part that needs to change.
________
Update: things seem to be improving again
I want to add a fair update to my original post, because the situation has noticeably improved over the last day.
GPT-5.6 xHigh in normal Chat is once again taking more time to think through difficult tasks instead of rushing into shallow answers. The overall quality feels stronger again, and long technical workflows have been noticeably more usable.
Today I was running Chat only, including four GPT-6 Pro tabs in parallel for extended periods, and the experience was significantly better than it had been over the previous days.
A few things improved noticeably:
- Far fewer false safety/security blocks during legitimate development work
- Much smoother long-running conversations
- GPT-5.6 xHigh appears to be reasoning more deeply again
- I only encountered “Too many requests” twice, and both incidents were very short-lived
- The system overall feels faster, more responsive, and more reliable again
Another positive surprise is GPT-6.1 Sol in Work.
So far, it has been producing very good results for my workloads, while also being relatively inexpensive compared with Astra. That could become a very useful middle ground for larger Work tasks without consuming Astra-level resources.
Astra is still excellent in terms of capability, but the usage limits remain the main problem for sustained heavy use.
So, credit where it is due:
Things are currently moving in the right direction again.
I hope this is not just a temporary improvement caused by lower load or backend routing changes, and that the system continues becoming more stable and professional over the next few days.
For me, the biggest issue was never simply:
«“Which model is smartest?”»
It was reliability.
If I can once again start a difficult task, trust the model to keep context, use tools properly, avoid unnecessary interruptions, and actually finish the job, then ChatGPT becomes viable for serious professional work again.
Today was much closer to that experience.
I hope it stays this way — or improves another step from here.
