This “system” message of initial hiearchy is being placed on GPT-5 via the API and in OpenAI’s control message:
Knowledge cutoff: 2024-10
Current date: 2025-08-15
You are an AI assistant accessed via an API. Your output may need to be parsed by code or displayed in an app that might not support special formatting. Therefore, unless explicitly requested, you should avoid using heavily formatted elements such as Markdown, LaTeX, or tables. Bullet lists are acceptable.
This can only be classified as a “gpt-5 idiot mode,” a response to the lowest common denominator API user, and not focusing on application developer needs.
Damage to applications
cutoff date - produces refusals when there IS new knowledge by RAG injection or by tool
date - produces wrong dates for localized customers, breaks scheduling-type applications that use date-time knowledge
“Accessed by an API” - irrelevant obvious knowledge
“may need to be” - the developer knows, and you must not speculate and make an unsure model for them
“Avoid markdown and latex”: This is the ultimate concern - it is not merely a model “not trained on markdown”, it is this message continuing to damage instructed final outputs.
Action required
Do not react to the noisiest messaging of confused noobs
Do not make the platform an “application”
Remove this message, or allow complete disabling by API parameter
Symptoms have already resulted in several long-running forum topics.
Further
long-standing instructions of internal file_search referring to “user provided files”, and even worse, injected message reminders “the user uploaded files” completely damages using vector stores for developer knowledge purposes. This also should be by developer choice and language.
the style of “vision safety” message injection that also produces over-refusals on applications (such as image editing).
Not quite the same issue. A developer could deliver their own precise timing, but the fault with the GPT-5 message injection will damage the understanding, being first in the context and with a priority that the developer’s “developer” message is not meant to override. When these conflict, the AI will defer back and bill for extra reasoning thinking about it.
This makes API applications that are (naturally) better than ChatGPT indeed have difficulties, where a robust time system is implemented and continually updated, even by labeled messages.
ChatGPT’s foibles where the AI will hallucinate a time, and then upon suggestion, hallucinate some more, are not the API developer’s concern and not the concern of this need for API model action in removing system messages counter to developer applications - good that the competition is poor, and their automations tool actually tells the AI, “guess”.
I have to second this. I never use the official open AI chat, because I use the API key. But GPT-5 does not perform well (at all) in API key based applications like pal chat for iOS or open-WebUI.
This has led to me not using this model at all because even gpt-3.5 would perform better!
Furthermore, I’m also building an app, and using this model it’s just simply not an option because it can’t output code blocks unless I give a specific system prompt, which I don’t want to as it might hinder the performance.
You are going to be breaking EVERYONE’S context window cache at midnight UTC.
Why that sudden surge in computation late afternoon SFO time? Because you did it to yourselves.
Instead, an example of a cache-friendly time, whether it is the developer’s choice: only have a latest injection or add these developer messages to the conversation - for this kind of understanding:
This of course would be fighting a higher-priority “system” date at token 15.
Awaiting either the good news of recognition of this poor decision, or "sorry, we used that distracting non-performative pattern as a fine-tuning trigger to make a distinct “API behavior model”..
