I'm getting an error on the first reponse on my agent. pipeline-error-openai-400-bad-request-validation-failed I think the root cause on your end is that openai dropped support for reasoning with completions when you have tools. Works when I select it in the dropdown, but now when I make an ephemeral agent

Suggestion: The error likely stems from a tool/model compatibility issue with GPT-5.4 and ephemeral agents. To resolve this, review your assistant's tool configuration and ensure only supported tools are enabled for the GPT-5.4 model. For example, when creating your agent, explicitly specify only default tools or remove custom tools:

json
{
  "model": "gpt-5.4",
  "tools": ["default"]
}

Try updating your ephemeral agent configuration as above and test again. For more details, see the [Vapi tools documentation](https://docs.vapi.ai/assistants/tools). Source: - [Introduction to Tools](https://docs.vapi.ai/assistants/tools)

I need to give custom tools

Yeah, this is a tough one, especially since it works in the UI but breaks with ephemeral agents. You’re right, it’s usually caused by using reasoning models with tools, which OpenAI doesn’t support anymore, so the request fails validation. Best fix is to switch to a standard chat model when using tools, or remove tools if you need reasoning. Also worth checking the ephemeral agent payload that’s often where things go off. Happy to take a look with you, are you passing tools in the config, or seeing this even without them? @droyster

it's a conversational agent, reasoning is off. I have it working out of vapi. it's either a misconfiguration on my end or a vapi bug. i'm leaning towards bug

If reasoning is already off and it works outside VAPI, this does feel more like a VAPI payload issue than your setup. The 400 usually means something subtle (tools, schema, or model params) is still being sent in the ephemeral request. I’d compare the raw VAPI request with your working config that’s where mismatches usually show up. Could you confirm if any tool config or extra params are still being attached under the hood? @droyster

Hi, thanks for reaching out.

thanks @Shaunak 019d64ce-4dab-7000-a5a6-c64692af237c 019d64c7-1ca3-7000-89fc-39da14fda149

Hi, Thank you for the call IDs and for your patience! We've looked into this and confirmed it's a bug on our end related to how certain OpenAI models handle tool/function calling in ephemeral agents. We've flagged it to our engineering team and a fix is being worked on. In the meantime, we'd recommend switching to gpt-4.1 or another standard chat model when using custom tools, as these work reliably with function calling. Apologies for the inconvenience and thanks for your patience! Best regards, VAPI Support