Credit where it is due: ChatGPT feels usable again

ChatGPT
,
1

@OpenAI_Support

Credit where it is due: ChatGPT feels usable again

After being pretty critical recently, I want to post a positive update — because today the experience has genuinely improved a lot for me.

And when things get better, I think it is only fair to say that too.

GPT-5.6 xHigh in normal Chat feels stronger again

This was one of my biggest complaints before.

For a while, GPT-5.6 xHigh felt noticeably weaker and less reliable than before the Astra release. It rushed more, lost context more often, stopped tasks early, and generally felt much less dependable for serious project work.

Today, that changed noticeably.

It is taking more time to think again, reasoning feels deeper, and long technical conversations feel much more coherent.

I can finally give it larger tasks again without constantly feeling like I have to babysit every step.

GPT-6 Pro has also been much more usable today

I had been getting far too many false security/safety interruptions during completely legitimate development work, especially when working through remote development/RDC.

That was one of the most frustrating parts of the experience.

Today I had four GPT-6 Pro tabs running in parallel for extended periods, and almost everything worked normally.

Very few false blocks.

That alone is a massive improvement.

“Too many requests” has improved as well

I still saw it twice today, but both incidents were very short.

Compared with the recent experience, where rate limiting sometimes made the whole workflow feel unpredictable, today felt completely different.

The system simply felt smoother.

GPT-6.1 Sol High is the biggest positive surprise

This deserves special mention.

GPT-6.1 Sol High in Work is producing extremely strong results for me.

And the most surprising part is the cost.

It is brutally cheap compared with the amount of useful work it gets done.

For my kind of workload — coding, infrastructure, long technical tasks, project work, and file-heavy workflows — this currently feels like one of the best price/performance points in the entire product.

Astra is still excellent. I still think Astra is incredibly capable, even at lower settings.

But Astra can consume usage very quickly.

6.1 Sol High feels like the first option where I can say:

This is strong enough for serious work, fast enough, reliable enough, and cheap enough to actually use all day.

That is exactly what I was missing.

I also received a full usage reset today

On top of that, I have now received a full reset of my usage limits.

Considering how much I have been using the system and how much additional capacity I had already purchased recently, this was hugely appreciated.

Thank you.

Seriously.

It immediately removed a lot of the pressure I had been feeling around constantly watching usage, credits, resets, and whether I could afford to let a long task continue.

Today finally felt professional again

My biggest frustration was never simply that one model was slightly smarter or slightly worse than another.

The real issue was reliability.

I need to be able to start a difficult job, let the system work, trust it to keep context, use tools correctly, avoid unnecessary interruptions, and actually finish what I asked it to do.

For a while, that trust was disappearing.

Today, it came back.

Today was the first time in a while where I genuinely felt:

“Okay. I can actually work normally again.”

That matters much more to me than benchmarks.

I use ChatGPT heavily and professionally. I am happy to pay for compute when the system gives me real value in return.

And right now, especially with 6.1 Sol High in Work, I am seeing that value again.

So: thank you to the teams working on this

There is a lot of negativity in the community right now, and some of it is understandable. I was one of the people complaining loudly myself.

But if I complain when things are bad, I also want to say thank you when they improve.

Today felt significantly better.

5.6 xHigh was stronger again.

6 Pro was much more usable for me.

Rate limiting was far less disruptive.

6.1 Sol High did excellent work at a surprisingly low cost.

And the full reset was very much appreciated.

Please keep moving in this direction.

If the system stays this reliable — or improves another step from here — then I am very happy to keep building my projects here.

For the first time in a while:

I can finally just work again.

Thank you.

2

Addendum: a concrete example from today

One concrete example of what I mean by the improvement today:

Right now I am having GPT-6.1 Sol High in Work handle a major architectural refactor of one of my projects.

This is not a small coding task.

The entire codebase structure is being reorganized and rewritten into a much more modular architecture. The main chat is acting as the QA orchestrator, while six subagents are working in parallel on different parts of the refactor.

At the time of writing, this run has been going for 95 minutes continuously.

I reset my usage limit immediately before starting.

And after 95 minutes of sustained work, with:

- GPT-6.1 Sol High

- a large codebase refactor

- six parallel subagents

- continuous QA/orchestration

- ongoing code analysis and restructuring

my remaining weekly limit still shows:

100%.

Seriously — wow.

This is exactly the kind of experience I was hoping for.

Strong model quality, long-running autonomous work, proper parallelization, and a usage cost that actually makes sustained professional work realistic.

If this is the direction OpenAI is taking with 6.1 Sol, then this changes everything for my workflow.

This is what I mean when I say I can finally work again.

So yes: this is getting somewhere, guys. Please keep it like this.

3

And what about “ChatGPT stream recovery polling timed out” and “Encountered exception: <class ‘caas.internal.errors.ClientError’>” are these errors fixed?
Using chat gpt in Medium was a torture.

4

lmao

Meanwhile, ChatGPT ignores ALL my instructions and repeats the same mistake every 2 minutes. I’m completely losing it here; this thing went from Einstein to toast-level intelligence within two weeks. I’m definitely canceling my subscription and switching to Claude. I’ve really had enough now, if you want anything more than “What’s the weather like today?” from ChatGPT, you’re completely screwed.

5

Unfortunately, that has not been my experience at all.

As of September 30, I’m still seeing the same severe regression across multiple unrelated chats and projects. Context is being lost, explicit instructions are forgotten almost immediately, corrections often don’t stick, and the model can repeat the exact same mistake right after being told not to do it.

I use ChatGPT heavily for several long-running projects, so this is not a minor inconvenience. Continuity of context is one of the biggest reasons I’ve stayed with OpenAI despite considering other providers.

I already have a huge amount of project history, decisions, workflows and accumulated context here. Moving all of that elsewhere would be a major undertaking, and that switching cost has been one of the strongest reasons for me to stay.

But if ChatGPT itself becomes unreliable at preserving and using that context, then OpenAI is effectively removing the main thing that keeps me here.

If I lose confidence in context continuity, there is very little left preventing me from moving my work to another provider. And ironically, the current regressions are making that migration more attractive.

So I don’t think this issue is resolved yet. At least for some heavy users, the reliability problems are still very much present today.

6

Same, and have been mentioning this for a long time. It’s not recent and many people have complained about this but no answer from any developers yet.

I’m fully with you and many here feel the same. We have also lost confidence in context continuity. Even several of my threads in ChatGPT suggested to try a different provider because it’s not able to keep continuity.

Local intelligence is high.
Longitudinal intelligence is unreliable - THIS IS STILL A HUUUUGE ISSUE.

I have also tried the obvious fallback: exporting my ChatGPT data and re-uploading the conversation JSON files.

That does not solve the problem.

OpenAI’s own documentation confirms why. Uploading an exported conversations.json file merely makes that data available as reference inside a new conversation. It does not recreate the original ChatGPT experience and does not restore memories, custom instructions, GPTs, account settings, original chat structure or other persistent state. OpenAI Help Center

This matters because one proposed workaround for long-context failure is often effectively:

“Export your data and start again.”

But that is not continuity.

A JSON archive containing the words of a conversation is not the same thing as restoring the working state of that conversation.

The relationships between decisions, corrections, instructions, rejected approaches, project state, memory and thread purpose still have to be reconstructed.

In my own case, even after exporting conversation history and re-uploading the JSON files, it did not recreate the accumulated working context I previously had.

So users currently have an awkward situation:

the live thread can lose effective access to older state, while the official export mechanism cannot restore that state as a functioning conversation either.

That is exactly why a proper continuity architecture matters.

This is bigger than one model giving worse answers.

Long conversations are losing reliable accumulated state. Memory is less transparent to the user than before. Exporting and re-uploading conversation history does not restore the original working state. And custom GPTs, another mechanism people used to create stable specialised environments, are now being retired in favour of plugins that OpenAI explicitly says may behave differently.

Each change may have an individual technical explanation.

Taken together, however, they are making persistent long-term use of ChatGPT harder rather than easier.

For users who have spent months or years building workflows inside this product, that is a serious regression.

7

Another example from today, and this one is even more concerning.

While using Codex, I have a dedicated app-building thread for one project. It has now started confusing that project with a completely different project, pulling in state that does not belong there.

So this is no longer just “the model forgot part of a long thread.”

This is cross-project state bleed / task-identity loss.

A supposedly dedicated project thread should not reach a point where I have to police whether it still knows which project it is working on.

This is exactly the wider problem I was trying to describe in the original post: local answers can still sound intelligent while the underlying longitudinal state is wrong.

I’ve now seen the same family of failure across writing, financial workflows, long-running project development, and app-building. Different domains, same pattern: established state gets lost, corrected behaviour returns, or one project starts contaminating another.

At this point this is clearly not an isolated bad conversation. It is a reliability problem with long-term state preservation.

Please investigate cross-project context contamination as well as ordinary context loss. A system that forgets what project it is in is not dependable for serious long-running work.