I’ve got gpt-5.6-Luna connected to my voice PE with web search enabled. No matter what I do, it seems like something upstream, probably on the model end, is overriding the current date and time. I can expose a date entity, I can add the current date and time into a prompt, but at some point in the evening the model decides it’s tomorrow.
For example, if I ask it at noon on Tuesday what’s the weather tomorrow, it will give me Wednesday’s forecast. But if I ask it at 10:30 PM on Tuesday what’s the weather tomorrow, it will give me Thursday’s forecast. Same behavior if I ask a question like did the Phillies win today.
If I ask it straight up what’s the current date and time, it gives me the correct answer.
I’ve tried to probe it a little bit and it seems to say there is an instruction with higher priority that tells it it’s the next day. My suspicion is that the model is being provided some kind of system clock which is set to UTC. I also have a suspicion that this is only invoked when Web search is used.
Has anyone else encountered or worked around this?
The system prompt (before your user prompt in that box) provides the datetime and yes I think UTC… And a get datetime tool the method, how it acts… Very much depends on which integration you’re using to talk to OpenAI. I see you’re using Luna - what integration are you using, Luna suppoort is supposed to come out in the 2026.8 build this week. Running beta?
There’s also some things that the ai doesn’t do very well unless time is provided with the tool request. (I’ve had problems in particular with certain oai models and time math… Just saying.)
For that we’d have to look at your voice debug logs to see what’s going on.
Yes, you should see if the date / time is in the system prompt when you look in the debug view right after the conversation.
You can also try to prompt in some help, like that you’re in a different time zone and it’s own date and time is incorrect. So it should always use the GetDateTime tool whenever current date and time is needed.
Call it a rule that always has to be followed and that it is not allowed to guess the date on it’s own, as it will be incorrect.
The model should normally be big and smart enough to follow that.
As a larger LLM, it should also be able to tell you where the wrong date is coming from.
Don’t simply ask like the end user of the voice assistant.
Tell it that you develop the assistant and try to improve the pompt and need to debug the problem.
This will often result to more precise answers.