I was picking a model for a voice note feature. The user records audio, we send it to Gemini, and get back JSON with a list of items. One call, no separate speech-to-text step.
I switched from 3.1-flash-lite to 3.5-flash-lite and expected it to get a bit faster. It got slower instead, so I measured it properly.
Same prompt, same responseSchema, same m4a file sent as inlineData, thinkingLevel MINIMAL, @google/genai 1.48.0, non-streaming, calling from SE Asia. 10 runs each, average:
| Model | text only | 12s audio clip |
|---|---|---|
| gemini-3.1-flash-lite | 1.43s | 2.68s |
| gemini-3.5-flash-lite | 1.20s | 3.88s |
| gemini-3.5-flash | 1.67s | 1.70s |
Look at the text column. With the same task as plain text, 3.5-flash-lite is the fastest one. So it is not a slow model. The problem only shows up with audio.
3.5-flash is almost the same speed with or without audio, so I don’t think this is a 3.5 problem in general.
I checked the easy answers first. thoughtsTokenCount is 0 in every run, both with MINIMAL and with no thinkingConfig, so thinking is not the cause. Every run returned 264-270 output tokens, so no model is writing more than the others.
The gap also gets bigger with longer audio. I sent the same file 2 times in one request:
| audio length | 3.1-flash-lite | 3.5-flash-lite |
|---|---|---|
| 12s | 2.56s | 3.88s |
| 25s | 3.76s | 6.81s |
That is about 0.08s per second of audio on 3.1, and 0.24s on 3.5. Our real clips are around 30s, which gives 4s vs 7-8s. Same as what we see in production.
So is this normal? And is there anything I can try to make audio faster on 3.5-flash-lite? Files API instead of inline base64, a different audio format or sample rate, media resolution?
Two small things I noticed, maybe they mean nothing. Pricing has a separate audio rate for 3.1-flash-lite ($0.50/1M for audio, $0.25/1M for text), but 3.5-flash-lite is a flat $0.30/1M for everything. Also, the 3.1-flash-lite model page shows a transcription example, and the 3.5-flash-lite page does not mention audio transcription at all. Does that mean 3.1 is the one meant for audio input?
Thank you!!