Gemini 3.8 Live is generally available, and the Live API's defaults moved with it

Google moved two audio-to-audio models to general availability on Tuesday: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both are reachable now through the Live API in the Gemini API and Google AI Studio, and the release notes put the GA on September 15 under the model strings gemini-3.8-live and gemini-3.8-live-extended-thinking. The model pages already list both as stable rather than preview.

The announcement is pitched at voice agents, and it lands five days after OpenAI put its own full-duplex voice model into general availability. But if you already run a Live API session, the migration notes matter more than the benchmark table.

What actually shipped

The model card says both models are built on Gemini 3 Pro. They take text, images, audio and video in, with an input window of 131,072 tokens, and emit audio and text up to 65,536 tokens. Google says they detect and transition between 97 supported languages mid-conversation, and that all generated audio carries a SynthID watermark.

The split between the two is about where the reasoning happens. 3.8 Live is positioned as the default for low-latency dialogue, with interleaved reasoning and asynchronous function calling out of the box. Extended Thinking reasons in the background while it keeps talking, using verbal acknowledgements and progress narration to cover the gap, and it exposes a thinking_level of low, medium or high. On 3.8 Live, that field is not accepted at all.

Beyond the API, Google says 3.8 Live is rolling out to Search Live for everyone and into Gemini Enterprise in private preview, while Extended Thinking reaches Gemini Live and Workspace surfaces including Docs for AI Pro and Ultra subscribers and Gmail and Keep for Google AI subscribers.

The scores are Google's own

Google reports that Extended Thinking takes the top overall spot on Artificial Analysis' Speech to Speech Quality Index with 82.6, scores 68.6% on the tau-Voice agentic benchmark and 35.1% on Sierra's tau-Voice-banking, and reaches 97.7% on Big Bench Audio. It places 3.8 Live second in the Speech Agent Arena. The company also cites ServiceNow's EVA-Bench, with a footnote that the run happened on the Live API on the Gemini Enterprise Agent Platform rather than the developer API.

All of those figures are Google's. Artificial Analysis is a third party, but its leaderboard was not reachable from our environment, so we could not open the ranking Google cites, and no independent evaluation of either model had been published at the time of writing. Speech-to-speech rankings are also sensitive to configuration: a competitor's score moves depending on whether it ran at a low, medium or high reasoning setting. Treat the ordering as a vendor claim until someone else runs it.

What it costs

The pricing page lists one rate card covering 3.8 Live, 3.8 Live Extended Thinking and the older 3.1 Flash Live preview together: $0.75 per million text input tokens, $4.50 per million text output tokens, $3.00 per million audio input tokens, $12.00 per million audio output tokens, and $1.00 per million image or video input tokens. Output pricing is explicitly described as including thinking tokens.

That single table is the point worth sitting with. Extended Thinking is not priced above 3.8 Live per token. It costs more because it emits more, and because its background reasoning lands on the output meter. A session that thinks harder shows up as a larger bill rather than a higher rate, which makes the thinking_level setting a cost control as much as a quality one.

The developer post quotes $0.005 per minute for audio input and $0.018 per minute for audio output, and its own footnote says those are estimates derived from the $3 and $12 token rates. They are not a billing unit. The actual meter runs on tokens, which matters because the Live API's default turn coverage now sends all video frames to the model. Google's documentation says so plainly and advises sending frames only when needed to manage context and cost. If your client streams a camera by default, that default is a line item.

The comparison with OpenAI is less clean than either rate card suggests. OpenAI's changelog put gpt-live-1 into general availability on September 10 at $0.05 per minute of voice session, billed per second, with backend model and tool usage charged separately. Google charges for the tokens and folds the reasoning into the same meter; OpenAI charges a flat session fee and bills the model doing the thinking on top. A per-minute number from one and a per-token number from the other do not compare directly, and anyone claiming a clean price winner is skipping the part where the workloads differ.

The defaults that moved

This is the section to read before scheduling the upgrade. Google's migration notes for moving off gemini-3.1-flash-live-preview list changes that will break working code rather than merely degrade it:

• Asynchronous, non-blocking function calling is now the default execution mode. Synchronous blocking still works on 3.8 Live if you set it explicitly on your tool declarations.
• thinking_level is not supported on 3.8 Live at all. Omit it, and omit thinking_config, from session setup.
• Proactive audio is permanently enabled. Setting proactive_audio to false returns an error.
• Affective dialogue has been removed from the API. Any enable_affective_dialog configuration has to come out of your code.
• Turn coverage defaults to including audio activity and all video, so video frames reach the model unless you manage them.
• Audio is the supported response modality. If you need a text transcript, enable output audio transcription.

Extended Thinking adds a protocol change on top of that, and it is the one most likely to produce a subtle bug rather than a loud failure. Because the model reasons in the background, turnComplete set to true no longer means the model is idle; the server may still be reasoning or waiting on tool calls. Clients have to keep listening after that signal and read the interaction_status field instead, which reports IN_PROGRESS or IDLE. Extended Thinking also supports asynchronous function calling only: blocking mode returns a hard error, and function scheduling configurations are not supported.

What the Live API still doesn't do

Both model pages list the same set of gaps: no context caching, no structured outputs, no code execution, no batch API, no URL context. Search grounding is supported. The missing structured outputs is the one that shapes architecture rather than cost. A voice agent that has to hand a clean JSON object to a downstream system cannot get it from the response schema here; the structure has to come out of a function call, or out of a second text model reading the transcript. That is a normal design, but it is a design decision the API is making for you.

The older gemini-3.1-flash-live-preview is now labelled a legacy Live API preview model in Google's own model list, with an explicit recommendation to update. Google's dedicated speech-to-text model, Gemini 3.5 Transcribe, was released last month and is not new today, though the developer post bundles it into the same audio suite.

The practical read: starting fresh this week, 3.8 Live is the reasonable default and Extended Thinking is for calls that genuinely involve multi-step work. With one already in production, budget the migration as a code change rather than a version bump, check whether your client treats turnComplete as an idle signal, and look at what your video stream costs before you look at the leaderboard.