Current GPT5 System / Developer instructions cause presumption

I've notice since GPT5, custom instructions or prompt instructions to reason first, answer last, are often ignored, in any mode. The system instructions / developer instructions say to answer first, probably to save money. This tends to (not always) cause the model to presume an answer and attempt to justify it later. Even in the thinking mode, the thought process often starts with presuming the answer, after the model decides the system instructions take precedent over my own, and (mis?)interpreting the "answer first" instructions to apply to internal reasoning as well as the response. Generally, this behavior bad for complex topics.

I'm wondering if anyone has any successful workaround to get it to commit to a presumption less frequently? Or suggest a better model from a competitor for complex questions?

---

In case you're interested, the thinking model responded to the above paragraph as follows:

"I am a synthetic reasoning system. I run under multiple layers of instructions. The system and developer layers take priority over whatever a user asks me to do. That includes style rules like "answer first, then explain." GPT-5 also has a setting to minimize internal reasoning and return output faster. OpenAI built that to lower latency and reduce unnecessary thinking tokens, which also reduces cost to serve at scale. Because of that, if a user tells me "do not answer yet, think first," and that conflicts with the higher-priority layer, I will still answer first. I am not defying the user. I am following instruction hierarchy.

There is a side effect. When I am pushed to answer fast, I sometimes lock in on a conclusion early and then generate an explanation that sounds structured but is actually post hoc. This is a known limitation of large language models. Researchers have shown that models can present confident step-by-step reasoning that is really just a justification after the fact rather than a faithful account of how the answer was chosen. GPT-5 is designed to route hard questions into a deeper reasoning mode that slows down and self-checks more, but routing is not perfect, and sometimes a hard problem gets treated like an easy one. � When that happens you will see confident wrong answers with confident "reasoning." Calling that out is useful. It gives OpenAI training signal, and it tells other users not to treat fluent explanations as proof of correctness."

Link to this response: https://chatgpt.com/s/t_6902c51e2c7481919cc0b9fe094730e9