Assistant API — instructions Parameter Confuses Model Even with Simple Prompts

ChatGPT Bugs
,
1

I’m running into an odd issue with the Assistant API, specifically when using the instructions parameter in client.beta.threads.runs.create. Even when I pass very minimal and clear extra instructions (e.g., the current time or a language setting), the assistant’s behavior becomes inconsistent or confused.

run = client.beta.threads.runs.create(
    thread_id=thread_id,
    assistant_id=assistant_id,
    model="gpt-4o",
    instructions="The current time is 2025-06-20 11:12. Please respond in Swedish.",
    tools=[{"type": "file_search"}]
)

With the above, the assistant either:

  • Fails to respond correctly,
  • Ignores the instruction altogether,

But when I omit the instructions argument entirely, the assistant works as expected.

My questions:

  1. Is this a known issue — that instructions may override/conflict with the system prompt too aggressively?
  2. Is this model-related? (e.g., more sensitive in gpt-4o than others?)
  3. Are there best practices for injecting minor dynamic context (like timestamps) without breaking assistant behavior?
  4. Has anyone found reliable workarounds?