Gemini 3.8 Live and Extended Thinking: API access, model choice and migration

Gemini 3.8 Live is GA, but the choice is about operating behaviour

Google made Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking generally available through the Gemini API on 15 September 2026. They are stable audio-to-audio models for the Gemini Live API, which maintains a stateful WebSocket connection for real-time exchanges. The launch matters less as a new pair of names than as a choice between two different ways a voice application can behave while work is under way. The base model is intended for prompt conversational response. The Extended Thinking model keeps reasoning and asynchronous tool work going in the background while it streams audio.

Google calls gemini-3.8-live its default for most low-latency voice-agent and real-time dialogue uses without reasoning-induced delays. It calls gemini-3.8-live-extended-thinking its high-reasoning option for complex, multi-step problem solving in a live voice interaction. Those are Google’s product positions, not a finding that one model is best for every call centre, assistant or embedded device. A team should decide from its own response-time target, task complexity, tool reliability and failure-handling design.

The practical model choice: fast dialogue or background reasoning

The standard model supports interleaved reasoning and asynchronous function calling, but Google’s documentation frames it around fluid, low-latency dialogue. That makes it the more straightforward starting point when the dominant job is turn-taking, visual grounding or quick voice responses and a delayed tool result can be managed separately. Its stable identifier is exactly gemini-3.8-live; it is not an alias for the older preview endpoint.

Extended Thinking is the option for a conversation that must continue while the service reasons through a multi-step task or waits for a tool. Google says it can reason and speak simultaneously, using background reasoning with asynchronous tool calls. The trade-off is architectural, not merely cosmetic: the client must stay ready for later audio frames or tool calls after a response appears complete. That is useful where uninterrupted acknowledgement and progress are important, but it can add state-management complexity and should not be selected solely because its name sounds more capable.

What both stable API models accept—and what they do not offer

Google’s model pages list text, images, audio and video as supported inputs for both models, with text and audio outputs. Each has a 131,072-token input limit and a 65,536-token output limit. The Live API overview describes raw 16-bit, 16 kHz little-endian PCM audio input, raw 24 kHz little-endian PCM output, JPEG images at no more than one frame per second, and WebSockets as the protocol. Developers handling visual streams should treat those transport details as part of the compatibility test rather than assume that a model-page modality label is a complete media-integration specification.

The same capability tables set useful boundaries. Neither model supports caching, code execution, file search, image generation, structured outputs, URL context or the Batch API. Both support Live API, function calling and Google Search grounding; only the standard model page describes function calling generally, while Extended Thinking says it is asynchronous only. This makes the models suitable for a real-time layer around carefully supplied tools, not a substitute for a separate file-retrieval, code-execution or structured-output workflow.

Background tool work needs explicit client and permission design

Asynchronous function calling means a declared function can execute without blocking the conversation. Google says Gemini 3.8 Live can acknowledge a request and keep talking while tools or API calls finish in the background. In Extended Thinking, that pattern is mandatory: behavior: NON_BLOCKING is the only supported function mode. Blocking mode returns a hard error, and function-scheduling configurations are not supported. The difference is material for teams that reuse tool definitions or client logic built around waiting for one synchronous result.

A background call is a scheduling feature, not an authority model. An application still chooses which functions exist, what data they receive, which credentials they hold, which actions need confirmation, and how a late, failed or duplicate result is reconciled. For browser-connected Live API clients, Google recommends ephemeral tokens rather than standard API keys in production.

Extended Thinking changes what a completed turn means

The most consequential integration detail is the session lifecycle. Google says that with Extended Thinking, turnComplete: true no longer proves that the server is idle. The service may still be reasoning in the background, awaiting an asynchronous tool response or preparing later audio. Clients must keep listening after that signal and use interaction_status on server messages. IN_PROGRESS means processing, background reasoning or pending asynchronous tools may continue; IDLE means the service has finished its current work and is waiting for input.

Extended Thinking also accepts thinking_config with thinking_level set to low, medium or high; MINIMAL is not supported. Its documentation says sending client content with turn_complete=true immediately interrupts active generation, while proactive audio is permanently enabled and setting proactive_audio: false produces an error. These are not edge cases to leave for production traffic. They affect how an interface labels progress, how an orchestration layer resolves a task, and when a user’s interruption cancels an answer.

A migration test should therefore include a tool result that arrives after turnComplete, an interrupted answer, a failed tool call and a duplicate or delayed response. The client should show whether work is still active, avoid declaring success before IDLE, and prevent a late message from being attached to the wrong user turn. These checks test the state contract Google documents without assuming that a successful demonstration covers every network or application condition.

Migrating from Gemini 3.1 Flash Live requires a configuration audit

For the ordinary low-latency path, Google’s migration guide says to replace gemini-3.1-flash-live-preview with gemini-3.8-live. It also says thinking_level and thinking_config are not supported for Gemini 3.8 Live, so they must be removed from session setup. Non-blocking asynchronous function calling is now the default; teams that require legacy blocking behaviour can explicitly set behavior: BLOCKING on their tool declarations for the standard model. Function scheduling options—SILENT, WHEN_IDLE and INTERRUPTED—are supported there.

Other changes can break a superficially successful upgrade. Google says full-session send_client_content updates now use explicit user or model roles; turn_complete=true interrupts active generation; affective-dialogue configuration has been removed; and the base model’s default turn coverage includes audio activity and all video. That means video frames are sent by default, so a product should only send them when context justifies the cost. The response modality is audio, with output transcription available when an app needs text. This is a reason to test actual session traffic rather than changing only a model string.

API access is direct; the wider product rollout is divided by surface

For developers, Google says both models are rolling out in the Gemini API and Google AI Studio. For enterprises, it describes private preview in Gemini Enterprise and says Gemini Enterprise for Customer Experience is coming soon. That wording is not interchangeable with general availability in every Google Cloud account. The announcement separately says Extended Thinking is coming soon to Google Workspace business customers, so an organisation should not infer that an API key, an AI Studio test or an existing Workspace subscription grants the same feature.

Consumer-facing surfaces are also distinct. Google says Gemini 3.8 Live is rolling out in Search Live. It says Extended Thinking is rolling out in Gemini Live; in Workspace Docs for Google AI Pro and Ultra subscribers; and in Gmail and Keep for all Google AI subscribers. The reviewed sources do not publish a universal country, language, plan or account eligibility table for these surfaces. Google’s launch post says the base model can transition between 97 languages in a conversation, while the general Live API overview lists 70 supported languages. That documentation difference is a reason to validate the exact product, locale and feature instead of turning either number into a broad availability promise.

Pricing and benchmark results are inputs to a test plan, not promises

At the source-check time, Google’s Standard paid-tier table gives the three listed Live models the same published rates: $0.75 per million text-input tokens, $3.00 per million audio-input tokens or $0.005 per minute, $1.00 per million image or video input tokens or $0.002 per minute, $4.50 per million text-output tokens, and $12.00 per million audio-output tokens or $0.018 per minute. Google lists a free tier as free of charge and says its paid tier is not used to improve products. These published rates are not a whole-deployment quote: they do not establish a country’s access, capacity, rate limits, Search grounding charges or the cost of an application’s own tools and media infrastructure.

Google’s launch post highlights an 82.6 Speech to Speech Index for Extended Thinking, 68.6% on tau-Voice and 97.7% on Big Bench Audio. Artificial Analysis, an independent benchmark publisher, listed the same 82.6 index, 68.6% agentic-performance and 98% speech-reasoning figures at review time, alongside methodology for its composite score. This is useful external corroboration that those results appear on the publisher’s current leaderboard. It is still bounded evidence: the index combines named tests, a particular Extended Thinking high setting and a tested tool environment. It cannot predict error rates, latency, safety or cost in an application with different prompts, networks, tools or users.

Safety claims and known limits remain part of the deployment decision

Google says audio generated by its AI products is watermarked with SynthID, and its model card describes internal safety evaluations and red teaming. The card also says the models can hallucinate, may encounter occasional slowness or timeouts, and are still subject to ongoing work on jailbreak resistance. It reports a January 2025 knowledge cutoff. These disclosures are more useful than treating a live spoken answer as inherently current or correct: time-sensitive information and high-consequence actions still need source checks, constrained tools and human review where appropriate.

The model card’s frontier-safety conclusion is also Google’s own assessment. Google says it found no meaningful new capabilities or material performance increases over Gemini 3.7 Flash that would likely reach its tracked or critical capability levels. That is not an independent certification of every implementation.

The sensible release-day conclusion is narrow. Gemini 3.8 Live gives a stable, low-latency route for real-time multimodal dialogue, while Extended Thinking is a different client-integration contract for work that continues after an apparent turn boundary. Test the model and configuration with representative conversations; measure tool completion, interruption handling, latency and spend; and check current Google documentation for the precise account and product surface. Neither the GA label nor a benchmark score establishes universal availability, compatibility or performance.