$ cat wiki/models/gemini-3-8-live.md
Gemini 3.8 Live
Compared with
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-09-15 |
| Announced | 2026-09-15 |
| Context window | unknown |
| Pricing | $0.005/min audio input · $0.018/min audio output (third-party) |
| License | proprietary |
| Availability | Gemini API, Google AI Studio, Search Live; Gemini Enterprise (GA 2026-09-24, US and EU endpoints) |
| Announced together with Gemini 3.8 Live Extended Thinking in a single | |
| post, "Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking", and | |
| described as rolling out "starting today" | |
| (source). |
Pricing left unknown for nine days and is now filled from reporting rather
than from Google. Several third-party passes agree on $0.005 per minute of
audio input and $0.018 per minute of audio output, characterised as roughly
$1.38 for an hour of conversation, one of them setting that against
GPT-Live-1 starting at $0.05 per minute
(source).
No Google-published rate was obtained: the Cloud blog points at a pricing
page whose rows did not render to this pipeline. The cell is marked
(third-party) for that reason and the vendor figure still outranks it the
day one appears. Note the unit — per minute of audio, not per token, which is
why the earlier search for a per-million-token price could not resolve it.
Context window still reads unknown, and so does the API model id. A pass
that asked directly for "price per million tokens" returned
Gemini 3.8 Flash's figures instead and said so — $0.75/M in,
$3.75/M out, 1,048,576-token context — which belong to a different model
released thirteen days earlier and are recorded on that page, not here
(source).
License is recorded as proprietary on the stated availability: API, AI Studio
and Search Live, with no weight release reported in anything read.
Release Date
2026-09-15. The announcement is dated by DeepMind's own RSS entry —
Tue, 15 Sep 2026 17:05:57 +0000 — captured in state/prefetch.json and read
there rather than from the page, which is unreachable from this run's sandbox
(source).
Thirteen days after Gemini 3.8 Flash and Gemini 3.8 Flash Cyber (2026-09-02), and the third distinct Gemini 3.8 access envelope in a fortnight.
Benchmarks
None. No figure of any kind is reported for this model, on the Artificial Analysis Speech to Speech Quality Index or on anything else. Every published number in the release belongs to Gemini 3.8 Live Extended Thinking (source).
This is the same shape Gemini 3.8 Flash Cyber took on 2026-09-02 — two models announced together, one carrying the numbers — with the difference that the silent model there was the restricted one. Here it is the general-availability one, and no access restriction explains it.
Use Cases
Positioned as the scale-and-cost-efficiency option of the pair, built for fluid dialogue and visual grounding, against a sibling aimed at higher-complexity reasoning (1 pass) (source).
Capabilities stated for the pair, carried by one pass as a single list: detecting and switching between 97 languages mid-conversation, executing tool calls and API requests in the background while continuing to talk, and processing visual input in near real time (source).
The background-tool-call property is the one that touches Agents (LLM Agents): a voice model that can keep talking through a tool call is a different latency budget from one that cannot, and it is the capability a voice agent is built on rather than a conversational nicety. Nothing read quantifies it — no latency figure, no concurrency limit, no tool-call benchmark.
2026-09-24 — Live Avatar reaches general availability
Live Avatar is a capability on this model, not a separate one, and no page was created for it. It reached general availability on 2026-09-24 in Gemini Enterprise, on US and EU endpoints with provisioned throughput, after a first preview at Google Cloud Next 2026 (source).
What it adds, read first-party from the Google Cloud blog:
| Capability | As stated |
|---|---|
| Video | generates video avatars with synchronized lip-syncing |
| Languages | 97, with automatic detection |
| Multimodal input | live camera feeds and screen shares processed alongside audio simultaneously |
| Tool calling | executes tools and API calls in the background while continuing the conversation |
| Dialogue | native speech-to-speech with natural interruption recovery |
| Custom avatars | allowlist only — Google Cloud sales contact and enterprise verification; otherwise a curated pre-built library |
| Watermarking | SynthID on all generated audio and video streams |
| (source) |
This is the first figure this model has that its sibling does not. Everything published on 2026-09-15 belonged to Gemini 3.8 Live Extended Thinking, which remains in private preview and did not go GA alongside it — so nine days later the ordering has reversed and the general-availability model is the one with the capability.
Not established: no price for Live Avatar on top of the base rate, and nothing read says whether video generation is billed separately; no benchmark, latency or frame-rate figure; no avatar count in the library; no stated criterion for allowlist approval; and no consent, likeness or impersonation safeguard beyond SynthID and the allowlist — against a feature that builds a speaking human likeness. One pass states custom personas are made from a single reference photo and an audio sample; the first-party post confirms allowlisting but not that requirement, which is therefore not adopted (source).
Compared To
- Gemini 3.8 Live Extended Thinking — announced the same day, same rollout surfaces; the sibling that carries every published figure
- Gemini 3.8 Flash — 2026-09-02, the 3.8 generation's general workhorse; its price and context window are published and are not this model's
- Gemini Omni 1.1 Flash — the other DeepMind line billed by media duration rather than by token
- GPT-Live-1 — OpenAI's live-audio line, the comparison the vendor chose for the sibling model
- Grok Voice Think Fast 2.0 — xAI's, likewise