Hello, I am facing extremely high latency with anthropic models. Upto 7-10 seconds repeatedly for first and subsequent messages. When i try to switch to a Google model, I get the Google llm failed message.

Pls switch to gpt if that is okay @d9232 let me know if u still face the issue

right now 40 cluster is whats working, but I really wanted to use gemini 3 flash as thats when I see best quality for my usecase long term. Any idea when this could be fixed as I have been facing this for a while on certain models.

Hi D9232, Can you send me the latest call_id? Along with that, can you provide me the duration in the recording where you had this issue?

This was the latest one where I faced issues with anthropic haiku 4.5: 019c4d5b-48cf-7dd5-99ed-8dffd81c774c. Assistant was supposed to speak first but they only spoke at around 20 seconds. IT seems like its not happening any more with anthropic. For gemini issues, here is a sample call: 019c512c-0b34-7ccc-a7d5-2095875e9e9a. It starts and fails with the llm failed error. Seems to be happening when assistant speaks first with model generated message and not in other first message modes

Gemini - LLM Failed Error Bug Found • Call ID:

019c512c-0b34-7ccc-a7d5-2095875e9e9a

• Error:

"contents are required"

from Google GenAI SDK • Root Cause: When

firstMessageMode = assistant-speaks-first-with-model-generated-message

, conversation starts with only a system message.

messagesToGoogleFormatted()

filters out system messages (passed separately as

systemInstruction

), resulting in an empty

contents

array. The fallback that adds a

(call started)

message doesn't take effect before the SDK validates the empty array. • Customer Workaround • Use

assistant-speaks-first

mode with a static

firstMessage

instead of

assistant-speaks-first-with-model-generated-message

when using Gemini models.

The opening sounds very different (and weird) for voices when i use firstMessage mode as static.