SunsettingAI Audio

Gemini 3.1 Flash Live Preview, Gemini 2.5 Flash Native Audio Preview

by Google

Announced

—

Complete

2026-11-17

Refund window closes

—

Refund status

No refund applicable

The Live API section of Google's Gemini deprecations page lists two preview models with a November 17, 2026 shutdown date: `gemini-3.1-flash-live-preview` (released March 11, 2026) and `gemini-2.5-flash-native-audio-preview-12-2025` (released December 12, 2025). Both name the same replacement, `gemini-3.8-live`, a stable model released on September 15, 2026. The page does not give an announcement date for these rows, so none is stated here. For 3.1 Flash Live users the price does not change, because Google prices 3.1 Flash Live Preview and 3.8 Live on one shared row. What changes is the session setup: 3.8 Live rejects `thinking_level`, runs function calls asynchronously by default, forces proactive audio on, removes affective dialogue, and sends every video frame to the model by default.

Refund flow

1
Not applicable. These are pay-per-token Live API models with no pre-paid balance tied to the model string, so there is no refund category to open.
2
If you are billed for either model string after November 17, 2026, treat it as a billing error and raise it with Google Cloud billing support with the invoice line, rather than as a refund request.
3
Re-budget with the paid tier prices per 1M tokens, as read from Google's pricing page on October 15, 2026. gemini-3.1-flash-live-preview and gemini-3.8-live share one row: input $0.75 text, $3.00 audio, $1.00 image or video; output $4.50 text, $12.00 audio. gemini-2.5-flash-native-audio-preview-12-2025: input $0.50 text, $3.00 audio or video; output $2.00 text, $12.00 audio.
4
From 3.1 Flash Live the rate card is identical. From 2.5 native audio, audio input and audio output are unchanged, text input is 1.5x and text output 2.25x, and video input is 3x cheaper per token. Cheaper video per token does not mean a cheaper session: 3.8 Live sends all video frames by default, so measure a real session before and after.

Migration path

Recommended successor

Gemini 3.8 Live (gemini-3.8-live)

Why this is the closest fit

It is the replacement Google names on both rows. It takes text, image, audio and video input, returns audio, supports function calling and Google Search grounding, and Google describes it as the default option for most low-latency voice agents. It is a stable model string rather than a preview, so it is not on the deprecations page with a shutdown date.

What differs from the original

Google's migration notes on the 3.8 Live model page are written for apps moving from gemini-3.1-flash-live-preview. Thinking: thinking_level is not supported, so omit thinking_level or thinking_config from session setup. If you need configurable reasoning, Google points to gemini-3.8-live-extended-thinking, which accepts thinkingLevel low, medium or high. Function calling: async (behavior NON_BLOCKING) is now the default, and you set behavior BLOCKING on a tool declaration to keep the old synchronous behaviour. Client content: send_client_content works for the whole session, and turn_complete true always interrupts the model mid answer. Proactive audio is always on, and setting proactive_audio false returns an error. Affective dialogue is removed, so delete enable_affective_dialog from your config. Video: turn coverage defaults to TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO, so every frame you stream is sent. Output: audio is the supported response modality, with output transcription for text. Limits on the model page: 131,072 input tokens and 65,536 output tokens, with no structured outputs, no caching and no Batch API.

Alternatives by use case

You stream a camera or screen into a Live session to describe or check video as it plays

Only send frames when the model needs them, because 3.8 Live sends all video frames by default. Compare token counts on one real session on both models before November 17

Your Live agent calls tools and expects the answer to wait for the result

Set behavior BLOCKING on those tool declarations, or rework the flow for async calls, which are the default on 3.8 Live

You set thinking_level on 3.1 Flash Live

Remove it for gemini-3.8-live, or test gemini-3.8-live-extended-thinking if you relied on higher reasoning

You use enable_affective_dialog or turn proactive audio off

Remove both settings. Affective dialogue is gone on 3.8 Live and proactive_audio false returns an error, so check how your agent behaves when it chooses not to answer

What this tool meant

These two previews were the main way to build real-time voice and camera agents on the Gemini API in 2026. The move is unusual in one way: for 3.1 Flash Live users the price is identical, so the whole cost of migrating is behaviour. Tool calls stop waiting by default, the model can decide not to answer, and streamed video is all sent unless you trim it. Google's own Live API guide still documents affective dialogue and proactive audio as opt-in features at the time of writing, while the 3.8 Live page says one is removed and the other cannot be turned off, so follow the model page when the two disagree.