Hi OpenAI team and community,
I’ve documented a set of regressions I’ve encountered with GPT-4o when performing structured technical work — specifically tasks that require strict formatting, deterministic code edits, or minimal changes to source logic.
Under GPT-4-turbo, my workflows consistently succeeded by using clear directives such as “fix only,” “do not change structure,” or “output full script.” With GPT-4o, these same prompts now frequently result in:
Unwanted logic rewrites, even when told not to- Code structure being reordered or removed
- Output truncation despite clear full-script instructions
- Disregard for spacing, indentation, or comments critical to tool compatibility
I’ve submitted these issues to OpenAI support, and they asked for detailed examples. I’ve created this repo as a permanent reference:
GitHub Regression Report: GitHub - chapman4444/gpt4o-regression-report: Documenting prompt obedience failures, formatting drift, and logic regression in GPT-4o.
It includes:
- Three real-world prompt/output comparisons
- Specific breakdowns of expected vs actual behavior
- Suggestions for resolving or mitigating these regressions
What I’m Requesting:
- A toggle for Strict Mode, where GPT prioritizes structure-preserving obedience
- Reinstatement or availability of GPT-4-turbo for technical users who require deterministic code generation
- Acknowledgement that “creative” interpretation of explicit technical prompts is a regression for these use cases
I’m sharing this so others can contribute their own examples or request similar capabilities. If you’re affected by these regressions, feel free to chime in or use the issue template included in the repo.
Thanks,
Ryan (chapman4444)