GPT-6.1 Sol over-refuses allowed Work/Codex tasks — GPT-5.6 fallback is not a viable solution

ChatGPT Bugs
1

I’m creating a new topic because the recent discussion “GPT-6 & Astra are far too sanitized. Almost unusable” was automatically closed after 24 hours:

I’m experiencing the same problem, but in my case it is breaking actual production workflows in ChatGPT Work.

I use Work for website administration, development, SEO, metadata generation, moderation configuration, debugging, content analysis, and other multi-step production tasks. Some of the websites contain lawful adult material.

GPT-5.6 Sol was generally able to complete these workflows.

GPT-6 and especially GPT-6.1 Sol increasingly refuse, sanitize, reinterpret, or stop tasks simply because the source website contains sexual material — even when the requested task itself is ordinary technical or editorial work such as:

  • SEO titles and descriptions
  • metadata
  • CMS development
  • plugin development
  • debugging
  • moderation configuration
  • content classification
  • site administration
  • analysis of existing material

This is not about asking the model to ignore safety policies.

The problem is false positives: an allowed production task becomes impossible because the model appears to treat the entire domain or source material as prohibited.

There is also a second major problem with the suggested workaround of switching back to GPT-5.6 Sol.

GPT-5.6 Sol consumes considerably more of the available Work/Codex resources. So, in practice, a Pro user is left with this choice:

  • GPT-6.1 Sol: much better usage efficiency, but repeatedly refuses or breaks otherwise allowed workflows.
  • GPT-5.6 Sol: can complete the work much more reliably, but consumes the available allowance too quickly for sustained Work usage.

That is not a workable fallback.

GPT-6.1 Sol is specifically positioned as a model for frequent complex Work/Codex tasks. If it cannot reliably distinguish prohibited content from legitimate technical, SEO, administrative, and analytical work involving a website that happens to contain adult material, that defeats the purpose of the model for this use case.

At this point the issue is serious enough that I’m questioning the value of maintaining a Pro subscription.

I would much rather see the false-positive/refusal behavior fixed than receive higher theoretical benchmark performance from a model that stops the workflow.

2

Unfortunately, I still haven’t been able to continue my work with GPT-6.1 Sol. The same refusal behavior continues to break my production workflow, so the practical issue remains unresolved for me.

I waited and retried, but I still cannot rely on the new Sol version for the work I actually use ChatGPT for.

Because GPT-5.6 Sol consumes the available Work/Codex allowance too quickly to be a sustainable fallback, I am now forced to cancel my Pro subscription.

I would prefer to stay subscribed, but I cannot justify paying for Pro if the more usage-efficient model cannot reliably complete otherwise allowed production tasks.

3

I am seeing the same underlying problem beyond adult-content workflows. In ordinary conversational and analytical use, GPT-6.1 often does not engage the actual statement. It sanitizes, paraphrases, or reinterprets it into a safer but materially different claim, then answers that substitute instead.

GPT-5.6 has generally been better at preserving established context, inferring the dominant meaning, and responding substantively rather than forcing repeated corrections. But it is no longer a dependable fallback because the application repeatedly overrides my manual selection: I was given 6.1, lost access to it, moved to 5.6, and am now repeatedly pushed back into 6.1. My reasoning-effort setting also changes from Extra High to Medium without my action, sometimes five to ten times in a day.

That creates the same impossible choice described here: 6.1 is efficient but unreliable for the actual work, while 5.6 is more reliable but is either constrained or overridden. Benchmark performance does not compensate for a model that refuses, sanitizes, or answers a different premise.

Please fix 6.1’s false-positive reinterpretation behavior, preserve 5.6 as a selectable model, and stop overriding explicit model and reasoning-effort selections.