3.1 flash TTS preview streaming latency

Gemini 3.1 Flash TTS: sub-2s with AI Studio’s default key, 10-20s with my own — capacity/routing issue?

I know this is a preview model with known bugs (e.g. the “OTHER” early-termination when streaming). This post isn’t about that — I’m trying to pin down latency specifically, and whether what I’m seeing is expected right now or something fixable.

Setup: Tier 2 (pre-paid), nowhere near rate limits. Short sentence, simple prompt, single speaker.

What I’m seeing:

  • My own API key: 10-20s to first audio chunk.
  • AI Studio with the default/playground key: consistently sub-2s, same text.
  • Same AI Studio session, swapped to one of my own keys: latency jumps straight back to ~15s.

That last point is the one that convinced me this isn’t the model’s ceiling — same UI, same text, only the key changes, and latency moves by 10x+. So something about routing or capacity tied to the key/account looks like the actual bottleneck.

One more oddity: this doesn’t look like real incremental streaming on my end. I get one long silent wait (~15s), then a burst of audio chunks arriving within milliseconds of each other — consistent with the full utterance being generated server-side first, then streamed out all at once, rather than chunks arriving as they’re produced.

Two working theories, don’t know which (if either) is right:

  1. AI Studio’s own key gets routed to different/priority capacity that customer keys don’t get.
  2. Something account-specific — despite paid Tier 2, my traffic is landing behind free-tier or unpaid traffic somehow.

Would genuinely like to hear: what latency are others seeing, on what tier, and has anyone found a setup that meaningfully improves this? (If the answer is “use Flash Live instead” — already ruled that out, voice quality doesn’t work for my use case.)

Quick update on this - things have got significantly better, and even using my key now, we’re down to the sub-2-second time, so I’m guessing it was resource problems on Google’s end.

Still keen to hear other people’s experiences, though.

Another update from me. The significantly better latency improvement that I saw only lasted a day, and now I’m back to unusable latencies.

Another observation to add to this: I have written some code to do a bake off between the generate content stream and the new interactions API. I’m finding that the new interactions API method of calling this model is significantly slower.

I will post more on this as I investigate.