ChatGPT Is Breaking Across Multiple Systems

[emoemeny]

I am posting this because what initially looked like an isolated GPT-5.6 behavior issue has now turned into a broader reliability problem affecting multiple parts of ChatGPT on my paid account.

The problems started suddenly around August 31 / September 1, 2026.

I have already contacted OpenAI Support, completed the standard troubleshooting steps, and had the case escalated. Support later confirmed that there was a platform outage during the timeframe I reported and stated that the outage had been resolved.

However, the problems described below continued after the public outage was marked resolved.

This is why I do not believe this can be explained only by a browser, cache, network, VPN, or session problem.

1. GPT-5.6 Sol in regular Chat suddenly behaves like an Instant/fallback model

This was the first and most obvious regression.

Before this issue started, complex requests in regular Chat with GPT-5.6 Sol would visibly take more time, perform deeper analysis, use tools when necessary, and produce substantially more complete results.

Since around August 31 / September 1, the behavior changed suddenly.

Typical symptoms:

• GPT-5.6 Sol is selected in the UI.

• Complex tasks receive extremely fast responses.

• The response often appears to skip deeper reasoning entirely.

• Instructions that previously required substantial analysis are completed superficially.

• Details are ignored more frequently.

• Repeated prompting to analyze more carefully often does not restore the previous behavior.

• Tasks that previously took meaningful processing time can suddenly return almost instantly with much poorer results.

This feels very different from a normal variation in answer quality.

There are also multiple community reports describing GPT-5.6 Sol or High/Thinking sessions apparently resolving to lower-tier/Instant-like behavior, including reports where users inspected request metadata and claimed to see GPT-5.5-mini or similar fallback behavior.

I cannot independently verify OpenAI’s internal routing on my own account, so I am not claiming that every affected request is definitely being routed to Mini.

What I can say with confidence is that the behavioral change is immediate, substantial, reproducible, and started suddenly.

2. Regular Chat and Work behave very differently on the same account

This is one of the strongest indicators that something is wrong beyond simple prompt quality.

I tested similar tasks using the same account.

In regular Chat:

• reasoning quality is reduced,

• file handling is unreliable,

• PDF analysis can fail,

• code/file-generation workflows can fail,

• tool-heavy tasks often stop or behave superficially.

In Work mode:

• the same or similar tasks often work substantially better,

• files are processed more reliably,

• multi-step execution is more capable,

• document generation that failed in regular Chat has succeeded in Work.

So this is not simply “GPT-5.6 became worse everywhere.”

The product experience itself appears to be routing or executing these tasks differently.

This difference between regular Chat and Work has also been discussed by other community users.

3. PDF and file analysis became unreliable

Another major regression is file handling.

Before this problem started, I could upload PDFs and ask ChatGPT to inspect their contents.

After the regression, I began seeing behavior where ChatGPT appeared to know that a PDF/file was attached, but its answers suggested that it had not reliably processed the actual contents.

In some cases it appeared to infer or guess instead of grounding the answer in the document.

This is especially concerning because there may be no obvious upload error shown to the user.

From the UI perspective, the file appears to be present.

But downstream file processing may not be happening correctly.

Other users have reported very similar behavior involving attachments that exist in the conversation but fail to materialize correctly into the execution environment or /mnt/data.

There have also been community reports involving:

• PDFs

• ZIP files

• CSV/XLSX files

• source files

• multi-file uploads

• intermittent sandbox access failures

This makes the problem much broader than one particular file format.

4. Python / container / file generation is also unreliable

I also encountered failures while trying to perform very simple file-generation tasks.

For example, I used straightforward ReportLab code to create a PDF:

from reportlab.platypus import SimpleDocTemplate, Paragraph, Spacer
from reportlab.lib.styles import getSampleStyleSheet

path = "/mnt/data/Balcony_Design_Trend_Book_2027_Vol01_Rebuild_Start.pdf"

doc = SimpleDocTemplate(path)

styles = getSampleStyleSheet()
story = []

content = """
Balcony Design Trend Book 2027

Volume 01 — The New Outdoor Room
"""

for line in content.split("\n"):
    story.append(Paragraph(line, styles["BodyText"]))
    story.append(Spacer(1, 8))

doc.build(story)

path

This is not a complicated workload.

It does not require an external API or complex computation.

Yet regular Chat struggled with execution/file generation in ways that had not happened previously.

Similar reports in the community mention:

• /mnt/data files not being available,

• caas.internal.errors.ClientError,

• uploaded files not appearing inside the container,

• container/Python failures even before meaningful file processing begins.

Again, similar workflows have worked better for me in Work mode.

5. Deep Research is now also failing

More recently, Deep Research has become another affected feature.

The failure mode is especially concerning:

Deep Research starts as if it is performing research, but it does not actually appear to search the web.

Symptoms include:

• no visible web searches,

• no meaningful source retrieval activity,

• no citations,

• no external sources,

• a final “research” report that appears to be generated without actual research.

In other words, the Deep Research UI can appear to run while the core research/tool chain is apparently not executing.

This makes the feature effectively unusable for tasks that require current external information.

I have also seen other users report the same pattern: Deep Research spending several minutes running but ultimately showing zero searches and zero citations.

This is especially problematic if a Deep Research usage/credit is still consumed even when no actual source research occurred.

6. This started during a real platform incident — but did not end when the incident ended

OpenAI Support told me:

“At the timeline you mentioned, our system experienced a platform outage that caused higher-than-usual error rates. This issue has now been resolved.”

This confirms that there was a real platform-side incident at the time the behavior changed.

However, that explanation does not resolve the current problem.

The public outage was marked resolved, but the degraded behavior continued afterward.

Since then, OpenAI has also had additional incidents involving ChatGPT, Work, file uploads, Codex, Responses API latency, and other services.

So there has clearly been broader infrastructure instability during this period.

But the persistent regressions described here appear to extend beyond a single resolved outage.

7. OpenAI has also acknowledged temporary model/feature availability restrictions

OpenAI staff/moderation responses in related community threads have stated that:

• access to certain models or features may temporarily be limited,

• this can happen even with an active paid subscription,

• limitations can be based on account activity,

• OpenAI cannot provide details about those checks,

• access is automatically reassessed.

This creates an additional question:

Are some of these regressions caused by an account-level availability/routing system rather than only infrastructure failures?

If so, users need much better visibility into what is happening.

A paying user should not select GPT-5.6 Sol while unknowingly receiving dramatically reduced reasoning/tool capability without any indication in the interface.

8. Troubleshooting already performed

I have already tried the standard troubleshooting steps suggested by Support:

• completely new chats,

• hard refresh,

• signing out and back in,

• logging out all active sessions,

• incognito/private browsing,

• extensions disabled,

• content blockers disabled,

• VPN/proxy disabled,

• different browsers,

• different devices,

• different networks,

• Lockdown Mode confirmed OFF.

The behavior persisted.

There are also no consistent upload/download error banners explaining the file failures.

This is why repeating generic browser troubleshooting is no longer useful.

9. Support case

I contacted OpenAI Support and explained the issue in detail.

The initial AI-assisted support repeatedly suggested browser/session/network troubleshooting.

After I confirmed those steps had already been completed, the case was escalated to a support specialist.

A human support response later confirmed the platform outage mentioned above but again suggested local troubleshooting after stating that the outage was resolved.

I replied explaining that the issue continued after the outage and that regular Chat and Work still behaved differently.

At this point I am still waiting for a substantive backend/account-level explanation.

Meanwhile, my paid subscription time continues to be consumed while several of the features I subscribed to are unreliable or unusable.

10. Why I am opening this bug report

I am not posting this to complain about one bad answer.

There appears to be a cluster of regressions affecting:

1. GPT-5.6 Sol reasoning behavior

2. model/routing consistency

3. regular Chat execution

4. file/PDF access

5. /mnt/data / attachment materialization

6. Python/container execution

7. generated file workflows

8. long or multi-step tasks

9. Deep Research web search

10. Deep Research citations/source retrieval

Multiple community users have reported overlapping symptoms during the same period.

OpenAI has already shipped some remediation/fixes for attachment access and has acknowledged that model/feature access can be temporarily restricted, but the overall problem does not appear to be fully resolved.

Expected behavior

If I explicitly select GPT-5.6 Sol on a paid account:

• the requested model/reasoning experience should actually be delivered,

• if a fallback or restriction occurs, the UI should clearly disclose it,

• uploaded files should reliably be available to the analysis/execution environment,

• PDF analysis should be grounded in the actual file contents,

• Python/container execution should reliably access required files,

• Deep Research should actually perform searches and retrieve sources,

• Deep Research should not complete with zero research activity while presenting itself as a research result.

Actual behavior

Instead, I am intermittently seeing:

• extremely fast, shallow responses,

• reduced or absent reasoning,

• failures to execute complex tasks,

• files visible in the UI but unreliable during actual processing,

• Python/file-generation failures,

• differences between regular Chat and Work,

• Deep Research completing without searches or sources.

What I am asking OpenAI to investigate

Please investigate this as a broader backend reliability/routing issue rather than another browser troubleshooting case.

Specifically:

1. Is GPT-5.6 Sol currently being silently routed or degraded for some paid users?

2. Are account-activity availability checks reducing reasoning or tool access?

3. Is regular Chat using a different execution/routing path from Work that is currently degraded?

4. Is attachment materialization into /mnt/data still unreliable?

5. Are Python/container provisioning failures still occurring?

6. Is Deep Research currently experiencing failures where the research/search tool chain is not invoked?

7. Can users be informed when a selected model or feature is unavailable instead of silently receiving degraded behavior?

8. Can affected paid users receive restoration of lost usage/credits or subscription time when features were unusable?

If other users are experiencing the same symptoms, please add your platform, subscription tier, affected model, approximate date/time, and whether the same task behaves differently in Chat vs Work.

The more reproducible examples we collect, the easier it should be to distinguish local issues from backend regressions.

[eclipsedmind]

Thank you for the detailed description of the problem. You have dissected the problem in great detail. I have experienced exactly the same symptoms, which started 3-4 days ago. Support has given me exactly the same answers. Please update if you find out what is causing the problem, and hopefully got a solution.

[ojfd]

I have similar problem.

Paid ChatGPT Plus subscription, running on “High”.

I have several contiguous multi-part technical chats.

I have set of rules for assistants governing Engineering Operating, Chat Transcript Use and Continuation Chat Startup Checklist.

I use markup transcrips of previous chats to “refresh the memory” of assistant in every new chat, since on High I run out of messages pretty fast.

Despite all these precausions, in my case, since about 24-25 August, the quality of work deteriorated considerably. Rules are forgotten almost immediately, consulting transcript too. Previous knowledge is ignored. When confronted and asked, assistant admits that he/she failed to apply the rules. Two large-ish projects went into serious regression.

I even used another chat to audit the transcripts of affected chat. Here is one excerpt.

For me that is indication that ChatGPT is broken at the moment. I’ve put my work on hold for now.

After this second audit pass, I would describe the two days like this:
There is not evidence that most of the supplied machine code was intentionally
inert.
There is evidence of four different quality failures operating together:

1. Actual defective implementations
   The original deferred-event destination and at least one failed experimental setup
   really were wrong.
2. Experiments that never validly measured what was claimed
   Examples include self-contaminating debugger traces, an uninstalled watchpoint,
   the failed polling probe, and debugger expressions unsupported by the actual tool.
3. Valid experiments executed on a confounded modified baseline
   This is the important new finding. Several later “we have ruled out X” conclusions
   were made while two major popup diagnostics remained active.
4. Valid but increasingly low-leverage investigation
   Large static audits and deeper AppKit branches were technically legitimate but
   produced repeated exclusions instead of moving toward a repair.
   So I would now change the wording slightly from my earlier answer.
   The strongest evidence is not that you were repeatedly fed bogus code.
   It is that you were repeatedly fed a mixture of:
   real code, some broken diagnostics, some confounded diagnostics, and many
   technically defensible but poorly prioritized diagnostics—while the prose
   often presented the accumulated conclusions as cleaner and more decisive
   than the experiment history justified.
   That can produce an especially strong “led by the nose” feeling because there is always
   a technically plausible reason for the next step.
   And unlike a deliberate-sabotage theory, that explanation is directly supported by the
   record.

[emoemeny]

I’ve Been Having This Problem Since September 1st. I Emailed Support But Haven’t Received A Proper Answer Or Solution.

I’ve Followed All The Other Threads And Problems Here And On Reddit But Haven’t Found A Solution To Fix This Problem.

I Found Out About This Process Last Night And I Want To Share It With You, Maybe It Will Solve Your Problem Too.

Chatgpt Has A New Security Update That I’ll Call The Session Timing System. Now They Can Give It Any Name They Want, But I’ll Give It This Name So You Can Understand What I Mean.

Now How Does This Update Work? First Of All, You Should Know This, When You Buy A Pro Or Plus Subscription, You May Give Your Password To Some Of Your Friends Or Acquaintances So That They Can Use It For Free. Some Sites Abuse This Feature By Sharing An Account Among Several People At A Cheaper Price And Controlling Each Person’s Login With A Two-Step Code. This Led Openai To Implement This New System That Prevents This From Happening, And I’ll Explain How It Works Below.

When You Log In To Chatgpt From A New Device, A Timer Is Set For Your Current Session, And In That Session You Can Do All The Things You Were Doing Before In Regular Chat. Such As: Analysis, Program Creation, File Creation, And Long Think. Just Remember That Deep Research Is Currently Disabled In Regular Chat And Does Not Search Any Resources. When Your Session Ends, Chatgpt Turns Into A Text Generator And Generates Its Own Answers, And Turns Into A Mini Model. Here Are Two Ways:

1- If You Are Using Desktop Systems, Log Out Once And Clear All Caches Related To Openai And Chatgpt In Your Browser Settings. Then Log Back In And Continue Your Work. If This Method Does Not Work For You, There May Be A Problem With Your Browser Settings. So You Can Use Anonymous Modes Like Incognito In Chrome Browser And Log In And Use It Without Any Problems.

2- If You Want To Do Your Work Without Having To End The Current Session, You Can Do It From The “Work” Section And All The Work Will Be Done There Without Any Problems. Deep Research Is Currently Only Available In The Work Section. Just Keep In Mind That If You Use The Work Section, Be Sure To Check The Usage Section So That You Do Not Face Any Restrictions.

Do Not Use Beta Versions For Use In Mobile Applications. The Beta Version Had A Problem For Me And After Exiting The Beta And Exiting The Previous Session, My Problem Was Resolved And Today I Am Doing My Work Without Any Problems. Do Not Consider This Solution As The Final Solution. I Shared This With You. It May Work For You So That You Can Finish Your Pending Work Sooner Before Your Subscription Ends. If You Get An Answer, Be Sure To Let Me Know.

[mcmillanjoseph18]

I think they nerfed 5.6 for astra im using astra in work chat and its doing what 5.6 was doing before the issues it’s just a theory but maybe openai soft nerfed astra to make room for astra i dont know but I been testing astra in work and it is doing what 5.6 should be capable of doing so I don’t know if openai changed something in there infrastructure and servers to give more servers to astra

[izzathnisfer]

Yeah, I was experiencing almost the exact same issue, and it was seriously affecting my daily work on last couple of days

Even when I was using GPT-5.6 Sol on High for complex tasks, it would reply almost instantly, as if it wasn’t really doing the deeper reasoning the task required. Repeating or rephrasing the prompt didn’t help much either, it would often keep repeating the same approach or make the same mistakes instead of properly reconsidering the problem.

This was especially frustrating because these were workflows and types of tasks that had worked much better before. I also raised the issue with support, but I didn’t receive a proper response.

Today I tried the same kind of work again, and so far it seems to be behaving normally again. The responses are taking an appropriate amount of time for more complex tasks, and the overall quality seems much closer to what I was getting before.

But sometimes I still feel like it’s not 100% back to normal. I’m not entirely sure whether the quality of the work and responses is as consistent as it was before.

Is anyone else still experiencing quality issues in works and noticing inconsistent behavior?

[Mike8two4]

Before today, my experience had been remarkably positive. I was genuinely impressed by how capable Astra was and by how efficiently it handled complex development tasks. I even praised the fact that I could accomplish substantial work without consuming an unreasonable amount of my usage allowance.

For several days, I was able to move multiple projects forward simultaneously while maintaining a careful, structured workflow. The system felt powerful, coherent, and surprisingly reliable. I was seriously considering upgrading my subscription because it had become an important part of how I work.

Then, within a very short period today, everything changed. My remaining usage allowance appeared to be consumed within minutes, much faster than I had previously experienced. At approximately the same time, the quality and reliability of the entire workflow began to deteriorate. Tasks that had worked smoothly earlier in the day suddenly became difficult or impossible to complete.
I do not think this is simply a case of receiving a few weaker answers. What I experienced today looks more like a breakdown across several parts of the workflow.

I use ChatGPT as a coordinated development environment for several separate software projects. One conversation acts as the project lead: it tracks decisions, reviews evidence, and prepares tightly scoped instructions. Codex in ChatGPT Desktop then performs the implementation using Astra at x-high reasoning and returns a structured text report. Git checkpoints protect the code, while fresh transfer prompts allow a new conversation to continue without relying on an extremely long chat history.

Until today, this system worked surprisingly well across several simultaneous projects. Then multiple failures appeared at roughly the same time:

• Project boundaries stopped being respected. Information and instructions from Project 1 began appearing in Project 2 or Project 3.
• One conversation treated its own project as though it belonged to a different conversation and should be avoided.
• The assistant returned to historical problems that had already been resolved instead of continuing from the latest verified checkpoint.
• It confused Codex in ChatGPT Desktop with Codex CLI and PowerShell, despite several explicit corrections.
• Some responses described what should happen but did not actually perform the requested work or produce the expected deliverable.
• Files were presented as completed deliverables, but their download links failed.
• New conversations and detailed transfer prompts did not reliably restore the expected behavior.
• Git protected the repositories themselves, but it could not prevent the assistant from reasoning from the wrong project history or giving instructions for the wrong project.

These failures may point to more than one underlying problem. The mixing of unrelated projects could indicate a context-retrieval or conversation-state issue. Missing deliverables and failed downloads may be a separate problem involving Work tools or artifact handling. The unusually short and shallow responses could reflect model routing, reduced reasoning allocation, system load, or another temporary service degradation.

The timing also overlaps with the Astra rollout and the suspension of new Pro 20X upgrades. That makes capacity pressure a reasonable hypothesis, but it remains only a hypothesis. I have no direct evidence that Astra’s computing requirements caused these specific failures.

What is clear is the practical impact: reliability degraded enough that continuing development became risky. For a casual question, an inconsistent answer is merely inconvenient. In software development, confusing Project 1 with Project 2, reopening obsolete tasks, or applying instructions to the wrong repository can damage working code, invalidate testing, or waste hours of carefully verified progress.

The most concerning part is not that the model occasionally makes mistakes. That is expected. The concern is that an established and previously reliable workflow deteriorated suddenly, and explicit corrections, Git checkpoints, isolated conversations, and detailed transfer prompts were not enough to restore it.

This should be investigated as a product-reliability issue rather than dismissed as normal model variability. The problem was not simply that the answers became slightly worse. The system temporarily lost the project separation, continuity, and execution reliability that had made it useful for serious development work

[krasnik]

This is a particularly useful report because Git protecting the repositories separates repository state from assistant/project state.

If the repo itself is still at a known-good checkpoint, but ChatGPT starts reasoning from Project 1 inside Project 2, revives already-closed tasks, or gives instructions for the wrong repository, that is not explained by ordinary code corruption. Something in context retrieval, project-state selection, conversation continuity, or orchestration is wrong upstream of the actual file edits.

I would keep the usage-collapse observation as a separate lane for now. The timing overlap is interesting, but it is not evidence that quota accounting caused the context failures.

There is another independent Work report from today that may be worth comparing: 9/12/26 ChatGPT Work 5.6 SOL Extra High Reasoning fails to follow instructions and "Work State" is missing . That user reports established Work context degrading, multi-change instructions causing collateral mutations while single-change instructions survive better, and one rendered final literally containing [object Object]. I am not claiming these have one cause, but they are useful same-day state/execution symptoms.

For your project-crossing case, the strongest next evidence would be:

• exact timestamp + timezone for the first wrong-project response;
• screenshot/export showing the wrong project name or resolved historical task appearing in the wrong conversation;
• whether the contamination appeared only in prose or also caused actual tool/file actions;
• whether a fresh conversation with a clean transfer prompt still selected the wrong project history;
• chat export / HAR if you can preserve one occurrence without exposing private source code.

The most important distinction is: the model making a normal mistake inside the correct project is one thing; the product supplying or selecting the wrong project state is a completely different class of reliability failure.

Your report is strong precisely because the Git checkpoints tell us the codebase can remain sane while the assistant’s idea of which project/history it is operating on becomes unsane.

[jerrymklu]

Same here, Chat mode is completely unusable now. Done a complete re-signin.

6 Pro or 5.6-Sol Extra High still produces answers instantaneously even when the task is large or complicated, and asking it to do something often produces acknowledgement replies which is completely useless. It no longer check project context, memory, web sources carefully, and just start guessing. Now all my replies start with “You are right…” because I have to keep correcting it every single time.

It does not reason anymore. It just outputs answers as bad as a language model does. It only searches a few sources online or quickly review memory if I explicitly ask it to. There are no multi-step thinking/reasoning iterations anymore, just a list of things it will do (recall project memory, search for something) and then just spits out a response that is such low quality, it is completely unusable. There is a ton of errors, all modes, including 6 Pro, is as dumb as Instant.

This has occurred maybe like once or twice in the past for about a day or so, where it does not reason at all, and just replies immediately with no further action, but they all recovered. This time it has been like this for days, and I really can’t stand this anymore. Kinda feel like the Pro subscription is getting scammed, and it not reasoning anymore probably saves the company a lot of money, but I really really want to believe that this is a genuine bug and can be resolved as soon as possible. I don’t know if this is trying to push us into using Work or Codex, but I’m not even trying to ask it to write code or perform tasks for me. Pro reasoning is part of the Pro subscription isn’t it?

I see another topic got closed, about 5.6 reasoning models dropping to 5.5 mini due to “usage limits”. My subscription cycle just started, I didn’t know there was usage limits for reasoning other than the Pro model for Pro subscriptions, but even if it did, the limit must have been reduced by a lot because there’s no way I exceeded the limit and still have not get a reset after days of frustration with this fast but dumb model. I would like an explanation or a fix, not just saying it’s a usage limit without telling me when it resets and when I’m being throttled, just like the Pro reasoning mode. Otherwise, how can I judge? If it can’t even reason on a 100~200USD monthly subscription, that’s kinda unfortunate.

[kingspatula3]

Same, I’m on my the plus plan and it gives me a stopped thinking and than a thinking and than a black screen and than a shallow response.. I wonder when it’s gonna be fixed.

[Pansonic]

This is happening to me as well since Sept. 15th. 5.6 Sol thinking has no reasoning and thinks really fast acting like instant. The replys are dry and dull and not following any instructions at all. Hitting the retry under each message and hitting 5.6 Sol seems to bump it back but has to done for each message.

There is another thread that I’ve replied too that has the same issues on here as well.

[kingspatula3]

Same here, it gives shallow responses and doesn’t check files or do actual research and thinking, or cite the source, it just gives a bland response.. I think it has something to do with ChatGPT tryna get rid of 5.5 on October 14 so the update is heavily infected on everything else, I wonder when this is all gonna get fixed.

[silverheart2195]

For me it takes too long to think on one step, or rushes through a thing and outputs the exact same result as the last test. It’s also incredibly slow right now and currently not doing anything while it’s thinking. How in the world has it gotten this bad???

[mcmillanjoseph18]

Possible False “Suspicious Activity Detected” Flag + “Log Out of All Devices” Failing — Anyone Else Seeing This? - #37 by mcmillanjoseph18 I just shared a update on 5.6 and it seriously feels like it has regressed it feels like it has been nerfed to make astra look good but im a pro member and astra isnt even available in normal chats or even project chats its available in work chats but not everyone wants to use work chat especially when astra burns through your usage like crazy , maybe openai should fix the issues first with the platform before releasing a new model and calling it a major improvement and I cant say for sure if 5.6 was nerfed to make astra look good but that is what it feels like .

[emoemeny]

Hello everyone. I’ve read through everyone’s messages, but I have a question: did any of you actually try the solution I posted? I shared a fix here, yet no one reacted; everyone just stated their problem and left. When I checked the thread again today to see if anyone had managed to resolve their issue using that solution, I was surprised to find that absolutely no one had even mentioned it. So, I asked myself: why should I bother sharing solutions for other problems here when no one takes the time to read them?
Do you realize that the issues you’re discussing are actually the most trivial ones?
Do you have a solution for when plugins freeze while working in the “Work” environment? Do you have a fix for the errors that occur when requesting image generation via GPT Image 2.5? And so many other issues… I ended up wasting my Plus subscription trying to find solutions for these problems. So, good luck finding fixes for the new bugs.

[toradrow777]

It did seem to work, the cache clearing thing. For me at least, for now. Thanks.

[benaya7]

Instead of going for weekly headline of new models, stabilize the runtime. I’m a plus user, but even in work mode I’m getting multiple errors. One is the " suspicious activity bug" you mentioned, otherwise execution sandbox unavailable in both modes, then it doesn’t load conversations at multiple occasions. Never happens with claude though. I believe they just push us gently towards the work mode, since it’s a bit more table and chat mode is not based on usage,but who knows.

[mcmillanjoseph18]

I agree instead of releasing new models so frequently fix the issues with the platform, fix normal chats and project chats ,fix all the issues first before releasing new features and new models

[Vyss]

This worked and restored full functionality for me instantly. Thank you!

[julesmarie86]

I have many of these issues. I went through support multiple times. They don’t care. I told them I canceled my sub because these issues have been going on since at least August. I’ve set aside certain projects because I can’t deal with all the course correcting and temporary fixes like clearing my cache (typically do every so often anyway). I won’t go into great detail other than to describe its performance as hot garbage.