Home → Help

The system prompt behaves differently across models

What you are seeing

• The model ignores instructions that worked on another provider
• invalid_request_error mentioning an unexpected role
• A system message appears to be treated as ordinary user text
• Behaviour changes when you switch models through the same gateway

Why it happens

Because both shapes are valid JSON for the other, mistakes rarely produce an error. The request succeeds and the instruction simply carries less weight, which looks like the model being worse rather than a wiring problem.

Multiple system messages are another divergence. Some implementations concatenate them, others keep only the first, and a few reject the request.

Reasoning models add a third variant: some treat developer instructions separately from user-visible system text, so where you put an instruction changes whether it survives into the reasoning step.

How to fix it

1. Keep one system message, firstOne system message at the start of the array works everywhere. Splitting instructions across several is where implementations diverge.
2. Test the instruction, not the response codeAsk the model to state its instruction back when you change providers. A 200 proves nothing about whether the instruction landed.
3. Avoid depending on system-only behaviour for hard rulesIf a constraint really matters, restate it in the user turn as well. It costs a few tokens and survives every translation layer.

Related

Last checked 2026-10-01. Written from problems diagnosed on a live OpenAI-compatible gateway, not collected from other sites.