Overview
Gemini 3.8 Live Extended Thinking is the high-reasoning half of Google's Gemini 3.8 Live pair, released to general availability on 15 September 2026 as the model id `gemini-3.8-live-extended-thinking`. Google's documentation recommends it over the default Live model when higher background reasoning is required for complex, multi-step problem solving during a real-time voice interaction.
The distinguishing feature is an asynchronous reasoning protocol: the model thinks in the background while it streams speech, which changes what the API's own signals mean. Google's documentation notes that `turnComplete: true` no longer indicates that the model is idle. Reasoning depth is set with `thinking_level`, which accepts low, medium or high here — `MINIMAL` is not supported. Proactive audio, which lets the model decide not to respond when input is not relevant to it, is permanently enabled on this model rather than opt-in.
Everything else it shares with its sibling. Google's model card states both 3.8 Live variants are based on Gemini 3 Pro, with a 128K-token input window (131,072 tokens), a 64K output limit (65,536 tokens) and a January 2025 knowledge cutoff. Text, images, audio and video go in; text and audio come out. The Live API caps audio-only sessions at 15 minutes and audio-plus-video sessions at 2 minutes, documents 97 languages the model can switch between mid-conversation, and lets any text-to-speech voice be selected through `speechConfig`.
One capability is narrower than on the default model: function calling supports only asynchronous, non-blocking execution — the synchronous blocking mode is unavailable. Pricing matches Gemini 3.8 Live exactly: $0.75 in and $4.50 out per million text tokens, $3.00 in and $12.00 out per million audio tokens ($0.005 and $0.018 per minute), and $1.00 per million image or video input tokens ($0.002 per minute).
| Released | 2026-09-15 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Context | 131,072 tokens |
| Max output | 65,536 tokens |
| Architecture | Based on Gemini 3 Pro |
| Knowledge cutoff | January 2025 |
| Modalities | Text, Image, Audio, Video |
| Status | Generally available |
Pricing
| Input | $0.75 / 1M tokens (text) |
|---|---|
| Output | $4.50 / 1M tokens (text) |
Audio: $3.00 in / $12.00 out per 1M tokens, or $0.005 / $0.018 per minute. Image and video input: $1.00 per 1M tokens, or $0.002 per minute. Identical to Gemini 3.8 Live, background reasoning included.
Strengths
- Background reasoning during a live audio session, so multi-step problem solving does not require ending the turn
- Configurable reasoning depth via `thinking_level` — low, medium or high
- Proactive audio permanently enabled: the model can decline to respond when input is not addressed to it
- Same audio-to-audio surface as Gemini 3.8 Live — 128K context, 97 languages, any TTS voice — at the same price
- Built on Gemini 3 Pro, per Google's Gemini 3.8 Audio model card
Best for
- Reach for it when a voice agent has to work a problem out during the call — triage, diagnosis, multi-step troubleshooting — rather than just answering.
- Reach for it for always-listening assistants where proactive audio should suppress replies to speech that was not aimed at the model.
- Reach for it when tool orchestration during a call is complex enough that low-latency-first behaviour produces worse answers than a short reasoning pass.
- Reach for it over Gemini 3.8 Live when answer quality mid-conversation matters more than the shortest possible time to first word.
How to access
| Provider | Model ID |
|---|---|
| Gemini API (Google AI for Developers) ↗ | gemini-3.8-live-extended-thinking |
Gemini Live (real-time audio) — every version
The full lineage of the Gemini Live (real-time audio) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Gemini 3.8 Livecurrent | 2026-09-15 | 128K | Proprietary |
| Gemini 3.8 Live Extended Thinking | 2026-09-15 | 128K | Proprietary |
FAQ
When should I use Gemini 3.8 Live Extended Thinking instead of Gemini 3.8 Live?
Google recommends it when higher background reasoning is required for complex, multi-step problem solving during real-time voice interactions. Gemini 3.8 Live is the default for low-latency dialogue without reasoning delays, so the trade is answer depth against time to first word. Both went GA on 15 September 2026 at the same price.
What does the asynchronous reasoning protocol change for developers?
The model reasons in the background while streaming its response, so `turnComplete: true` no longer means the model is idle — code that treated that flag as an end-of-work signal needs revisiting. Function calling is also restricted to asynchronous, non-blocking execution; the synchronous blocking mode is not available on this model.
How do I set the reasoning depth?
Through `thinking_level`, which accepts low, medium or high on this model. `MINIMAL` is not supported here. That knob is the main lever for trading spoken-response latency against how much reasoning happens behind the conversation.
What is proactive audio and can it be turned off?
Proactive audio lets the model decide not to respond when the incoming content is not relevant to it — useful for always-listening setups where not every utterance is addressed to the assistant. On Gemini 3.8 Live Extended Thinking it is permanently enabled rather than an opt-in flag.
What are the context and session limits?
Google's Gemini 3.8 Audio model card gives a 128K-token input window (131,072 tokens) and a 64K-token output limit (65,536 tokens). Separately, the Live API limits audio-only sessions to 15 minutes and audio-plus-video sessions to 2 minutes, so a long conversation is bounded by the session clock as well as the window.