Gemini 3.8 Live vs Extended Thinking: Pricing, Features and API Migration

What launched on September 15?

Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 as two generally available audio-to-audio models for the Gemini Live API. Google positions the standard model around responsive voice dialogue and the Extended Thinking model around difficult requests that require deeper reasoning, planning or several external tools. Google's release notes record both model IDs as generally available.

Developers can build with both in the Gemini API and Google AI Studio. Consumer and enterprise product rollouts are separate: Google's launch announcement says standard 3.8 Live is rolling out in Search Live, while Extended Thinking is rolling out in Gemini Live and particular paid Workspace experiences. Do not interpret an API model's general availability as a promise that the same model option is visible in every Gemini app, country, workspace or account.

Sources: Google's model announcement and Gemini API release notes.

Gemini 3.8 Live vs Extended Thinking at a glance

Both accept text, image, audio and video input and can produce audio responses. Google documents the same published input/output token limits on their model pages; the decisive differences are reasoning flow, tool execution and how the client knows an interaction is still active.

Sources: 3.8 Live model card, Extended Thinking model card and Live API Thinking guide.

What does “Extended Thinking” actually do during a voice call?

The standard model is designed to answer promptly. It can use tools, but its conversational turn generally concludes when turnComplete arrives. In contrast, Extended Thinking can issue an intermediate spoken update while continuing to reason or waiting for an asynchronous tool result in the background.

Imagine a travel assistant receiving, “Check flights, compare hotels and tell me whether I can arrive before dinner.” Extended Thinking could speak an update such as “I’m checking the flight options” while separate external searches run, then deliver a fuller answer later. The intermediate speech is not the final result. That difference is why Google added an explicit interaction-status signal for this model.

For a simple language-practice voice bot or an assistant that turns on a light, the lower-latency model may fit the interaction design better. For a voice agent that must coordinate multiple APIs and keep talking naturally while those operations run, Extended Thinking provides the additional background workflow described in Google's documentation.

Source: Google — Thinking in the Live API.

Gemini 3.8 Live and Extended Thinking API pricing

Google currently places both new 3.8 Live models, along with the older 3.1 Flash Live Preview, in the same Standard Live API pricing table. The following are published paid-tier US-dollar rates per one million tokens unless a minute equivalent is shown.

Google also lists a free API tier for this group of models. Free-tier availability does not mean every API integration or every duration is unlimited. The same pricing section says Google Search grounding has a shared monthly allowance of 5,000 free search requests across Gemini 3.x models on paid tier, followed by $14 per 1,000 requests. A single model request can trigger more than one search query.

These are API consumption rates, not a Gemini Live app subscription comparison. Google may change prices, tiers or grounding allowances, so check the live rate card before relying on a budget estimate.

Official rate card: Gemini Developer API pricing.

Why Extended Thinking can have a different total bill at the same token rate

A shared per-token rate is not the same as an identical per-request total. An Extended Thinking session might use more thinking/output tokens while analyzing complex input, make additional grounded searches or remain active longer. A quick standard Live exchange might finish with less generated output. Measure the actual token counts and tool usage for your application before estimating the difference.

Audio-to-text transcription can add a separate charge: Google's Live API best-practices guide says enabling input or output transcription causes generated transcript text tokens to be billed at the text-output rate in addition to the normal audio-token cost. Avoid assuming transcripts are included for free in the audio rate.

Cost-conscious teams should also review session duration, active listening, unnecessary video frames, model output length and whether a workflow truly needs high background-reasoning depth. For comparison, test the same task types against both models and use recorded usage, not just the headline price table.

Sources: API pricing and Live API cost and streaming best practices.

Where can developers and regular Gemini users access the models?

Developers: both 3.8 Live model IDs are generally available through the Gemini Live API and Google AI Studio. Use the model ID shown in the model cards when opening a Live session; these are real-time conversational endpoints, not simple drop-in replacements for ordinary text-only generateContent requests.

Consumers: Google's September 15 announcement describes 3.8 Live rolling out to Search Live and Extended Thinking rolling out to Gemini Live. The announcement separately lists private-preview or upcoming enterprise experiences and specific Google AI subscriber access in Workspace products. A user not seeing a model name in one app should not assume the API is unavailable, or vice versa.

Sources: Google's rollout announcement and Developer audio-model release.

How to migrate from Gemini 3.1 Flash Live Preview to 3.8 Live

If your voice application uses gemini-3.1-flash-live-preview, Google identifies gemini-3.8-live as its recommended replacement. For the standard 3.8 Live model, start with a small configuration migration and then regression-test your audio and tool workflows.

1. Replace the model string with gemini-3.8-live in the Live session setup.
2. Remove thinking_level and thinking_config from standard Live setup. Those controls are unsupported for the standard 3.8 Live model.
3. Review tool behavior. Async calls are the default. If your existing function depends on synchronous execution, explicitly declare behavior: "BLOCKING"; standard 3.8 Live still supports it.
4. Review new 3.8 settings. Proactive audio is permanently on. Remove a legacy proactive_audio: false flag and any old enable_affective_dialog setting.
5. Keep normal turn-completion handling for the standard model, then test barge-in, output audio, transcription and tool return paths using a real session.

Google's standard-model card says the supported response modality is audio; enable output-audio transcription when the interface needs text. That differs from merely requesting a text-only response for a voice-first session.

Source: Gemini 3.8 Live migration notes.

Migrating to Extended Thinking requires three client changes

Switching from standard Live or 3.1 Flash Live Preview to Extended Thinking is not only a model-name substitution. Google identifies three integration points that need explicit handling:

1. Track the overall interaction. Treat IN_PROGRESS as ongoing model work, even after one spoken utterance ends. Return the interface to listening/idle only after IDLE.
2. Declare every function as non-blocking. Use behavior: "NON_BLOCKING" for Extended Thinking tools; blocking declarations fail.
3. Set supported background-thinking depth only if needed. Use low, medium or high. Do not use minimal.

Extended Thinking also does not support function-scheduling configuration. Remove scheduling options such as those used with some standard Live functions before connecting the Thinking model.

Sources: Live API Thinking migration guide and Extended Thinking model migration notes.

Why turnComplete: true no longer means “all done”

In standard 3.8 Live, a completed model turn allows your application to return to an idle or listening state. Extended Thinking can emit multiple spoken updates during one user request, with further reasoning, audio or tool calls still to come.

In the wire response, the field can appear in camel case as interactionStatus; the Python SDK exposes interaction_status. If your frontend stops receiving events at the first turnComplete, it may truncate the final answer, ignore a later tool call or incorrectly enable a fresh microphone turn while the model is still busy.

Implementation rule: continue listening for server messages throughout an Extended Thinking interaction, and update the overall UI state using IN_PROGRESS and IDLE.

Source: Google's side-by-side Live protocol guide.

Extended Thinking requires asynchronous tool declarations

Standard Live accepts both synchronous blocking and asynchronous non-blocking functions. Extended Thinking accepts non-blocking functions only. Google says synchronous blocking tool mode produces a hard error when used with this model.

This distinction is separate from whether your own external service is quick or slow. Even a fast function needs the correct declaration when you use Extended Thinking. Your application still needs to execute the function, match the response to the pending call, handle errors and keep the session open until the model finishes its overall interaction.

Google's standard model supports function scheduling options, but its Extended Thinking model card says function-scheduling configurations are unsupported. Remove those configurations rather than carrying every old tool declaration across unexamined.

Sources: Standard model tool behavior and Extended Thinking tool behavior.

Model IDs and thinking configuration examples

For the raw Live setup message, Google shows the models/ prefix. If you use a Google GenAI SDK, pass the model ID in the shape supported by that SDK. The snippets below illustrate the different setup fields; they are not complete deployable applications or substitutes for session authentication and audio streaming.

The Extended Thinking example sets a supported reasoning level. If tools are included, each function declaration must additionally specify NON_BLOCKING. In SDK settings, the same concept may be represented as thinking_config with thinking_level="low".

Standard Gemini 3.8 Live does not accept thinking_level or thinking_config. Remove those fields rather than setting them to minimal or an empty value.

Source: Google's Live API setup and SDK examples.

Proactive audio can change the cost of a migrated session

Google says proactive audio is permanently enabled in both 3.8 Live models. Sending proactive_audio: false returns an error. The older gemini-3.1-flash-live-preview did not support that feature, and Google notes an associated billing distinction: for 3.8, input tokens are charged the entire time the API is listening, while output tokens are charged only when it responds.

If your app leaves a microphone stream connected for long stretches of silence, compare actual usage after upgrading rather than multiplying only the speech-active seconds by an approximate per-minute rate. Use your app's legitimate listening-state controls, input pipeline and session lifecycle to avoid unnecessary active audio time.

Google also says video frames are sent to standard 3.8 Live by default under its updated turn-coverage behavior, so avoid streaming frames your task does not need. For session history, the Live API documents context-window compression as a way to manage retained conversational history and related cost in long sessions.

Sources: Live API best practices and Standard 3.8 migration details.

Voice, visual context, languages and supported input

Google designed both releases for native real-time speech conversations, including visual grounding where the application streams image or video context. Its developer announcement highlights asynchronous function calling, accurate handling of codes and other alphanumeric data, and support for 97+ languages across the new Gemini Live models.

Both model cards list 131,072 input tokens and 65,536 output tokens. Google lists text, images, audio and video as input types and text/audio as output data types, while the standard-model migration note specifies audio as the supported Live response modality and recommends transcription for a text display. The supported output data types should not be mistaken for a promise that a normal real-time voice session can simply switch to a text-only response modality.

Neither model is the same endpoint as gemini-3.8-flash, the separate general-purpose Flash model. Use the Live API when the product needs bidirectional audio interaction and follow the relevant Gemini model card for text-generation or batch use cases.

Sources: Google's developer audio announcement, Live API capabilities guide, and both standard and Thinking model cards.

Is Gemini 3.1 Flash Live Preview being shut down?

Google's deprecations table currently lists gemini-3.8-live as the recommended replacement for gemini-3.1-flash-live-preview, but it shows no shutdown date announced for that older preview model as of this review. A recommended replacement is not the same as an already-enforced retirement.

Developers can migrate on their own testing schedule while monitoring Google's release notes and deprecations page for a later dated announcement. Preview model behavior or availability can still change, so do not leave critical applications dependent on an old preview indefinitely.

Source: Gemini model deprecations and replacement table.

Common Gemini 3.8 Live migration errors

A practical migration and launch checklist

1. Choose gemini-3.8-live for immediate voice dialogue or gemini-3.8-live-extended-thinking when the product truly needs background multi-step reasoning.
2. Confirm the Live API endpoint, account/API tier and the model ID shown in Google's current model cards.
3. For standard 3.8 Live, remove unsupported thinking configuration and review default asynchronous tools and permanently enabled proactive audio.
4. For Extended Thinking, mark tools NON_BLOCKING and handle IN_PROGRESS / IDLE separately from spoken turn completion.
5. Regression-test audio input/output, visual context, interruptions, function errors and long-running requests.
6. Measure token usage and active listening time with transcription and grounding enabled, if you use them.
7. Monitor the 3.1 Flash Live Preview deprecation listing rather than assuming its recommended replacement date is a shutdown date.