Gemini 3.8 Live arrives as two voice models — 30.1% and 68.6% on the same agentic test
Google's Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are generally available, and the company's own benchmark document shows the two voice endpoints trading conversational quality and cost against multi-step tool use.
Google released two real-time voice models to general availability on September 15, 2026:
gemini-3.8-live and gemini-3.8-live-extended-thinking. Both are stable audio-to-audio endpoints in the Gemini Live API that take text, images, audio and video and answer with speech and text, and both reach the Gemini app, Google AI Studio and Vertex AI — the extended endpoint also Gmail, Docs and Keep. They replace gemini-3.1-flash-live-preview, now labelled a legacy preview. 12What launched
| Signal | Confirmed detail | Action window |
|---|---|---|
| Gemini 3.8 Live | Default Live endpoint for low-latency voice agents, with interleaved reasoning that adds no delay to replies. 1 | If you run voice agents on the 3.1 preview, gemini-3.8-live is the new target string. |
| Gemini 3.8 Live Extended Thinking | For sessions that plan: background reasoning runs during the audio stream, tools execute asynchronously, and the model speaks fillers such as "Checking flights now" while it waits. Depth is set with thinking_level at low, medium or high. 3 | Use it where a tool takes seconds to answer, and measure the added latency before swapping it into an existing agent. |
| Asynchronous function calling by default | On gemini-3.8-live, behavior: NON_BLOCKING is now the default execution mode, with blocking kept for compatibility. On the extended endpoint, blocking tools return a hard error. 4 | Audit tool declarations first: a blocking-only tool works on one endpoint and fails on the other. |
Session state moves off turnComplete | During background reasoning, turnComplete: true no longer signals an idle session. Clients read interaction_status (IN_PROGRESS or IDLE) to know when the model has finished. 5 | A client that returns to "listening" on turnComplete will cut off a late answer. |
| Removals and new defaults | Affective dialogue is gone; proactive audio is permanently on and proactive_audio: false errors; audio is the only response modality; video frames now count in the turn by default. 36 | Delete the removed configuration, and send video frames only when the session uses them. |
The two endpoints are a price-and-reasoning trade
Google's evaluation document puts the two apart on the same benchmarks. Artificial Analysis's agentic τ-Voice test scores Extended Thinking at high reasoning 68.6% and standard 3.8 Live 30.1%; on its Speech-to-Speech index the gap narrows to 82.6% against 76.0%. Cost moves the opposite way: Google reports $0.84 per hour of input audio on 3.8 Live and $3.50 on Extended Thinking at high. Extended Thinking also leads Sierra's τ³-Banking leaderboard at 35.1%, ahead of GPT-Live-1 Astra at 32.0%. 7 Artificial Analysis, ServiceNow and Sierra ran those tests; Google published them.
The shared ceiling is the same on both: a 128K-token context window, 64K output, and no context caching, code execution, file search, structured outputs, URL context or Batch API. Google's model card lists hallucination, occasional slowness and timeouts as known limitations, and a January 2025 knowledge cutoff. Paid audio runs $3.00 per million input tokens and $12.00 per million output tokens, with a free tier alongside. 28
turnComplete, while the extended endpoint holds interaction_status at IN_PROGRESS through background reasoning and asynchronous tool calls. Source: Thinking in the Live API.Why it matters
Buyers of voice agents now choose between two endpoints from one vendor on a measured spread: the cheaper model wins on conversational experience, the expensive one wins on multi-step tool use by more than a factor of two. The decision rests on published numbers, so it is worth making deliberately. Two things to watch: a shutdown date for
gemini-3.1-flash-live-preview, and independent latency measurements, since Google's own results cover accuracy, experience and cost.References
- 1Gemini API release notes
ai.google.dev
- 2Gemini 3.8 Audio model card
deepmind.google
- 3Gemini 3.8 Live Extended Thinking
ai.google.dev
- 4Live API capabilities guide
ai.google.dev
- 5Thinking in the Live API
ai.google.dev
- 6Gemini 3.8 Live
ai.google.dev
- 7Gemini 3.8 Audio evaluation methodology and results
deepmind.google
- 8Gemini Developer API pricing
ai.google.dev
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.