Gemini 3.8 Live API: Pricing, Extended Thinking and Setup

Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026. They are native speech-to-speech models for the Live API: you talk, they talk, tools can run in the background, and the conversation does not have to go silent while a function returns.

The model IDs are gemini-3.8-live and gemini-3.8-live-extended-thinking. Paid audio is $0.005 per minute in and $0.018 per minute out. Context is 128K in and 64K out. They are not Gemini 3.8 Flash. Flash is chat. These are live voice.

They are not on RouterPlex yet. Both IDs land on this catalog in a few days, at Google list rates, 0% markup, on the same prepaid key as the rest of the lineup. Until then, Flash is the live Google route.

Two models, two jobs #

Google's own thinking guide is the useful split. Do not pick Extended Thinking because the name sounds smarter.

Both take text, images, audio, and video. Both return text and audio. Both support search grounding and function calling. Neither supports caching, batch, code execution, file search, structured outputs, Maps grounding, image generation, or URL context.

Proactive audio is on and cannot be turned off. Setting proactive_audio: false errors. Affective dialogue is gone: drop enable_affective_dialog if you still send it.

Gemini 3.8 Live API pricing #

Google bills both Live 3.8 IDs at the same paid rates. Output includes thinking tokens.

A full duplex hour where both sides speak the whole time is 60 × $0.005 + 60 × $0.018 = $1.38. Real calls have silence, so that is a ceiling, not a typical invoice. Thinking tokens still land on the output line. A high thinking session that talks little can still cost more than a chatty gemini-3.8-live turn.

Google's per-minute audio figures match 25 audio tokens per second: 1,500 tokens/min × $3 / 1M = $0.0045, published as $0.005; 1,500 × $12 / 1M = $0.018 exactly.

Search grounding is the Gemini 3.x bucket: 5,000 free requests per month, then $14 per 1,000. A customer prompt can fire more than one search.

When these IDs land on RouterPlex, the billed rate is the live catalog row, not a screenshot in this post. No per-token markup. Prepaid balance. Hard per-key budget. The session stops at $0. It does not overdraft.

Cost of one voice hour, in numbers #

Give a voice key its own budget. A forgotten open mic is a continuous input stream at $0.005/min, plus whatever the model says back, plus thinking. That is a different failure mode from a 40K chat completion.

Gemini 3.8 Live specifications #

Google's product post also claims alphanumeric precision (confirmation codes, claim numbers) and visual grounding: the model can use what the camera sees while you talk. Default turn coverage includes video frames. Only send frames you mean to pay for.

Benchmarks Google is citing #

These are Google's numbers from the 15 September announcement, plus the Artificial Analysis speech-to-speech table checked the same day. They are not RouterPlex evals.

Google also says the pair sits on ServiceNow's EVA-Bench Pareto frontier for accuracy vs conversational quality, run on the Live API on Gemini Enterprise Agent Platform.

Read the table as "this is a voice agent model," not "this replaces Flash for coding." τ-Voice for plain Live is 30.1% on AA. The Extended Thinking lift is the point of the second ID.

What changed from Gemini 3.1 Flash Live #

If you are on gemini-3.1-flash-live-preview, Google's migration notes:

• Change the model string to gemini-3.8-live.
• Omit thinking_level / thinking_config on the non-thinking ID.
• Async NON_BLOCKING tools are the default. BLOCKING still works on Live, not on Extended Thinking.
• send_client_content works for the whole session. turn_complete=true interrupts generation.
• Proactive audio is permanent.
• Response modality is audio. Turn on output transcription if you need a text log.

For Extended Thinking, the extra work is client state: keep listening after turnComplete, and treat interaction_status as the idle signal. Google's thinking doc is the protocol, not a blog gloss.

Gemini 3.8 Live vs Gemini 3.8 Flash #

A voice agent that needs 2026 facts should turn on search grounding, or keep a Flash/chat model on the same prepaid balance for lookup. The Live cutoff is older than Flash.

On Google's products today, on RouterPlex in a few days #

Google is rolling the models out now:

• Developers: Gemini API and Google AI Studio Live.
• 3.8 Live for everyone: Search Live.
• Extended Thinking for everyone: Gemini Live; Docs Live for Google AI Pro and Ultra; Gmail Live and Keep Live for Google AI subscribers.
• Enterprises: private preview in Gemini Enterprise.

Partners named for the Live API plumbing: Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, Vision Agents. Sample apps: google-gemini/gemini-live-api-examples.

RouterPlex: gemini-3.8-live and gemini-3.8-live-extended-thinking are scheduled for this catalog in a few days. They will not appear as a drop-in /v1/chat/completions model. They need a realtime session. Do not paste the ID into Cursor or Claude Code and expect a voice agent.

Until the rows exist, the Google chat model already on this key is gemini-3.8-flash. Create a prepaid account, set a hard budget, and use Flash for text. When the Live IDs publish, put them on a separate key with a small cap. An open microphone plus thinking tokens is how a voice demo becomes an invoice.

Create a RouterPlex account if you want the key ready. The Gemini 3.8 Flash pricing post is the current Google chat route. This page will follow the catalog the day the Live IDs ship.