Recurring Python Execution Failures, Locked Chat Bugs, Work Usage Limits, and GPT-6 Reliability Issues

I have been using ChatGPT for a long time for academic research, translation, document analysis, and other complex tasks.

Since November 2026, I have been experiencing several recurring issues that significantly affect productivity and the reliability of ChatGPT as a practical working tool.

I would like to share these experiences and find out whether other users are encountering similar problems.

1. Recurring Python and File Processing Failures

Since November 2026, the Python execution environment has repeatedly become unavailable during ongoing tasks.

The following error has occurred multiple times:

ClientError
Encountered exception: <class 'caas.internal.errors.ClientError'>.
  • Even simple Python commands such as print("Python OK") cannot be executed.

  • File reading, editing, generation, and integrity verification become impossible.

  • Creating a new conversation sometimes temporarily resolves the issue, but the same problem eventually occurs again.

  • The issue has affected multiple conversations.

  • The official OpenAI status page sometimes indicates that all systems are operational while these issues are occurring.

Moving to a new conversation requires preparing handover documents, uploading files again, and reestablishing working conditions. This creates a considerable amount of unnecessary work.

I would appreciate improvements to the stability of the execution environment, a way to recover it without abandoning an existing conversation, and more informative public status reporting.

2. Recurring Issues with Locked Chats

Since November 2026, I have experienced two different recurring problems with the Locked Chats feature.

Permanent unlocking failure:

When attempting to permanently unlock a conversation, an error appears stating that locked chats are temporarily unavailable and that I should try again.

Afterward, the same conversation appears in both the regular sidebar and the Locked Chats list.

The conversation remains protected by a PIN, but the permanent unlocking process does not complete successfully.

This issue occurs on both the desktop browser and smartphone app.

Repeated PIN authentication prompts:

Even after entering the correct PIN, the authentication screen repeatedly reappears, preventing access to the conversation.

Reloading the browser temporarily resolves the issue, but it has occurred multiple times since November 16, 2026.

Since this feature is intended to protect user privacy, its reliability and consistency are particularly important.

3. ChatGPT Work Usage Limits

I have reached the usage limit after submitting only two tasks in ChatGPT Work.

On another occasion, after resetting the usage allowance, I reached the limit again after approximately ten minutes.

I understand that resource consumption varies depending on task complexity and the model being used. However, the current situation makes it extremely difficult to predict how much work can actually be completed.

Previously, I was able to carry out similar tasks to a reasonable extent. Since the beginning of November 2026, however, I have been reaching the usage limit after submitting only two task instructions.

This noticeable change raises concerns about whether the usage limits or the way resource consumption is calculated have changed.

I would appreciate:

  • Greater transparency regarding usage consumption.

  • Visibility into the resources consumed by individual tasks.

  • Estimated resource consumption before starting a task.

  • Sufficient usage capacity to make Work practical for sustained tasks.

Introducing new capabilities is valuable, but those capabilities also need to be practically usable.

4. Decline in Practical Reliability Since the Introduction of GPT-6

Since switching to GPT-6, I have repeatedly experienced the following problems:

  • Insufficient document comprehension: The model fails to thoroughly read attached documents and overlooks explicitly documented requirements.

  • Poor instruction adherence: The model acknowledges instructions but fails to follow them during execution.

  • Unilateral reduction of task scope: The model independently reduces the requested workload without adequate justification.

  • Incomplete self-correction: The model acknowledges its mistakes but repeatedly apologizes or explains them instead of correcting the actual deliverables.

  • Excessive task fragmentation: Tasks are divided into unnecessarily small portions, requiring frequent instructions to continue.

  • Inconsistent decision-making: Identical working instructions and procedures produce inconsistent decisions across conversations and different stages of the same task.

  • Unnecessary interruptions: Work stops prematurely even when further meaningful progress appears possible.

These problems persist despite detailed instructions, documented workflows, quality control procedures, and comprehensive handover documents.

After switching back to GPT-5.6, I observed a noticeable improvement in working stability.

Although this is not a controlled scientific comparison, it suggests that the difficulties cannot necessarily be attributed solely to the user’s instructions or working methods.

I believe model evaluation should consider not only reasoning benchmarks but also practical performance indicators such as:

  • Instruction adherence.

  • Task completion rates.

  • Successful correction of identified errors.

  • Consistency throughout long-running tasks.

  • Frequency of unnecessary interruptions.

  • Required human intervention.

  • Total human time required to complete a task.

I sincerely hope that GPT-5.6 will remain available until these practical reliability issues with GPT-6 are sufficiently addressed.

Final Thoughts

I appreciate the continued development of ChatGPT and the introduction of new capabilities.

However, technological advancement should be accompanied by improvements in reliability, stability, and practical usability.

The purpose of using AI is to reduce human workload. Recently, I have found myself spending an increasing amount of time managing AI-related problems, repeating instructions, rebuilding working environments, and recovering interrupted tasks.

In my experience, this is becoming a significant obstacle to productivity.

I hope OpenAI will place greater emphasis on the stability of existing features and the practical reliability of its models, alongside the development of new technologies.

Are other users experiencing similar issues?

I would appreciate hearing about your experiences, and I hope this feedback will reach the relevant development and product teams.

Thank you.