Gemini Live API 2026: Models, Benchmarks & Gemini 4

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are the newest models on the Gemini Live API, released September 15, 2026, and marked Stable. This guide covers the developer API behind them. The consumer Gemini Live app has its own page, linked below.

Google's launch post for Gemini 3.8 Live reports vendor-measured benchmark results. For a comparison outside Google's own results, we ran Gemini 3.8 Live, 3.8 Live Extended Thinking, and 3.1 Flash Live Preview through 82 call scenarios, three runs each, in Cekura's speech-to-speech benchmark.

Across all 82 scenarios, including callers who change their mind mid-sentence, the older 3.1 Flash Live Preview reached 62.2% reliability, ahead of the two 3.8 models at 52.4% and 41.5%.

TL;DR: What is the Gemini Live API?

Google released Gemini 3.8 Live and 3.8 Live Extended Thinking on the API on September 15, 2026, and Cekura tested both.

• The Gemini Live API is Google's WebSocket API for live voice and vision agents. It takes streamed audio, images, and text, and returns spoken replies.

• Gemini 3.8 Live and 3.8 Live Extended Thinking are the newest models on it. Google calls 3.8 Live its default option for most low-latency voice agents, and Extended Thinking adds reasoning while it speaks. Both cost $3.00 per 1M audio input tokens and $12.00 per 1M audio output tokens.

• 3.1 Flash Live Preview passed all three runs on 62.2% of scenarios, 3.8 Live Extended Thinking on 52.4%, and 3.8 Live on 41.5%. Treat those figures as the live-call baseline and rerun them on your own scenarios.

What is the Gemini Live API?

The Gemini Live API is a stateful WebSocket API from Google for building low-latency voice and vision agents on Gemini models. Per Google's Live API docs, it accepts raw 16-bit PCM audio at 16 kHz, JPEG frames at up to 1 frame per second, and text, and it returns 16-bit PCM audio at 24 kHz.

The current models are gemini-3.8-live and gemini-3.8-live-extended-thinking, both released September 15, 2026, as listed on the deprecations page. Each accepts up to 131,072 input tokens and returns up to 65,536 output tokens, according to the 3.8 Live model page.

Key features of the Gemini Live API

The Live API treats a conversation as one stateful session, so most of its features concern what happens while a caller is speaking. Interruptions, tool calls, transcripts, and language switches all work mid-call:

• Barge-in: Callers can interrupt mid-reply, and the server sends a message so your client can stop and clear queued audio.

• Tool use: Function calling and Google Search grounding both run inside a live session, and grounding costs $14 per 1,000 requests after 5,000 free requests a month.

• Transcription: Input and output transcripts are separate settings, so you choose which side of the call you log. The 3.8 models respond with audio only, so turn on output transcription if you need text.

• Language switching: Google's launch post says 3.8 Live detects and switches between 97 languages mid-conversation.

• Thinking control: Extended Thinking accepts low, medium, or high, and 3.8 Live takes no thinking level.

Gemini Live API vs. the Gemini Live app: what's the difference?

The Gemini Live API is the developer interface, and Gemini Live is the consumer voice feature inside the Gemini app. Both run on the Live model family, but developers and consumers reach it through different surfaces and set different options.

If you're shipping an agent to callers, the API is where you set a fixed model ID and declare tools. If you want to talk to Gemini yourself, go to Google's Gemini Live page.

How does the Gemini Live API work?

The Gemini Live API works over a persistent WebSocket connection that streams audio in and audio out while the model reasons and calls tools.

1. Open a session: Your backend connects over WSS and sends a setup message with the model ID, system instructions, tools, and audio settings.

Alternatively, a browser or mobile client can connect directly, which Google says generally performs better for streaming. For that route, Google recommends ephemeral tokens, currently in Preview, over API keys.

2. Stream input: Send 16 kHz PCM audio, plus JPEG frames at up to 1 per second if the agent needs visual input. Google's best practices recommend 20 to 40 ms chunks for streaming, and small chunks of 20 to 100 ms in client buffering, to minimize latency.

3. Receive audio: The model streams back 24 kHz audio. Automatic voice activity detection is on by default, with a recommended silence duration of 500 to 800 ms.

4. Run tools in the background: On the 3.8 models, function calls default to NON_BLOCKING, which lets the conversation continue while a tool runs.

Without compression, audio-only sessions last 15 minutes and audio-video sessions last 2. A single connection lasts around 10 minutes.

Context window compression extends session length, and resumption tokens stay valid for 2 hours after a session ends, according to Google's session management docs.

What's new in the Gemini Live API?

The newest change is the 3.8 generation, released September 15, 2026. The table below lists the main Live models with their status from the deprecations page.

Shutdown dates are the earliest possible retirement dates on the deprecations page, checked October 29, 2026.

Google's launch post says 3.8 Live handles visual input with minimal delay and runs tools in the background during a conversation. Extended Thinking reasons while it speaks and describes its progress on multi-step tasks.

The 3.8 Live model page calls 3.8 Live the default option for most low-latency voice agent experiences and live dialogue.

Migration notes from 3.1 Flash Live Preview to 3.8 Live

The 3.8 Live model page documents these changes for anyone leaving gemini-3.1-flash-live-preview. The launch post doesn't mention any of them.

• Model string: Switch to gemini-3.8-live.

• Thinking: Remove thinking_level and thinking_config from session setup.

• Function calling: NON_BLOCKING is the default. Set BLOCKING on a declaration when a tool must finish before the model speaks.

• proactive_audio: It's on for every session, and setting proactive_audio: false returns an error.

• Affective dialog: The API removed it, so delete enable_affective_dialog.

• Video: Turn coverage defaults to TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO, so frames are included automatically. Send frames only when the agent needs them, to control context and cost.

• Client content: send_client_content works all session, and turn_complete=true interrupts active generation.

• Output: Responses are audio only, so turn on output transcription if you need text.

Extended Thinking supports only non-blocking function calling (BLOCKING returns an error), and turnComplete: true no longer means the model is idle, so check interaction_status for IN_PROGRESS or IDLE before treating a turn as finished.

Gemini 3.8 Live vs. Extended Thinking: which should you choose?

Google's model pages recommend 3.8 Live for low-latency dialogue and Extended Thinking for multi-step problem solving. The table below draws on Google’s Live API capabilities page and the two model pages.

In Cekura's benchmark, Extended Thinking scored 52.4% pass³ against 41.5% for 3.8 Live, and the extra reasoning added only 0.16s to its median response (2.07s against 1.91s).

Google's Extended Thinking model page documents the interaction_status handling covered above.

Where is Gemini 3.8 Live available?

Developers can use both 3.8 models in the Gemini API and Google AI Studio, while Gemini Enterprise and the consumer apps got them in stages. The September 15, 2026, launch post placed the 3.8 models like this:

• Gemini API and Google AI Studio: Available for the two 3.8 models from launch day.

• Gemini Enterprise: Private preview for both. Gemini Enterprise for Customer Experience was listed as coming soon.

• Consumer surfaces: 3.8 Live in Search Live, and Extended Thinking in Gemini Live.

Google added Gemini 3.8 Live with Live Avatar in Gemini Enterprise on September 24, 2026, and custom avatars require enterprise allowlisting.

How much does the Gemini Live API cost?

Gemini 3.8 Live, 3.8 Live Extended Thinking, and 3.1 Flash Live Preview share one paid rate card. Audio input costs $3.00 per 1M tokens, which Google puts at $0.005 per minute, and audio output costs $12.00 per 1M tokens, or $0.018 per minute.

Google describes the paid tier as prepaid and pay-as-you-go, and the table below breaks the rates out by input and output type.

Free-tier usage costs nothing, and Google uses that data to improve its products. Paid-tier data stays out of that use.

Grounding with Google Search includes 5,000 free requests a month across Gemini 3 and newer models, then costs $14 per 1,000 requests, with a default cap of 1,500 requests a day per project.

The older 2.5 Flash Native Audio preview charges the same $3.00 and $12.00 for audio, with text priced at $0.50 per 1M input tokens and $2.00 per 1M output tokens.

The benchmark leaderboard shows $0.070 per minute for 3.1 Flash Live Preview, averaged over test calls using Google's published rates, and leaves both 3.8 models blank.

The per-minute rates in the table cover audio and image tokens only, so compare them with the cost per call in your own usage logs.

Pricing verified against the pricing page on October 29, 2026. Confirm the current rates there.

For per-minute costs worked out turn by turn, and why long calls cost more, see our Gemini Live API pricing breakdown.

How does Gemini 3.8 Live perform on benchmarks?

In Cekura’s live-call benchmark, 3.8 Live finished 12th of 12 at 41.5% reliability and Extended Thinking 9th at 52.4%.

Google’s own tests report results mainly for Extended Thinking, which scores 82.6 on Artificial Analysis’ Speech-to-Speech Quality Index (ranked first) and 68.6% on τ-Voice.

Google also reports 35.1% on Sierra's τ-Voice-banking and 97.7% on Big Bench Audio. It says 3.8 Live placed second in Speech Agent Arena, with no score.

These are vendor-measured results, and Google shows its EVA-Bench results only as a chart.

Google's evaluation methodology says the Artificial Analysis and τ³-Bench runs used gemini-3.8-live-preview and gemini-3.8-live-extended-thinking at high thinking, as of September 2026. Any Artificial Analysis or τ-Bench results shown for 3.8 Live came from that preview ID.

The Cekura benchmark baseline

The benchmark ran 12 speech-to-speech models and a cascade reference through 82 scenarios, three times each, over live WebRTC calls (3,198 calls in total). The cascade reference runs speech-to-text, then an LLM, then text-to-speech, and the benchmark doesn't rank it.

Each model used the same Pipecat agent, prompt, and tools. A call passes only if the outcome is right and every saved field matches, and reliability (pass³) counts scenarios passed on all three runs.

The benchmark covers 3.1 Flash Live Preview, 3.8 Live Extended Thinking, and 3.8 Live. Cekura published it on September 23, 2026, and updated it on October 22, 2026.

We design and run the benchmark ourselves and sell testing tools, and some ranked vendors may be our customers or partners.

Its methods, test cases, agent code, and call recordings are public, but the Expected Outcome judge's prompt and some scoring rubrics aren't yet, so those scores can be re-run but not exactly reproduced.

3.1 Flash Live Preview had the best reliability of the three Gemini models. It passed all three runs on 51 of 82 scenarios. Extended Thinking reached 43, and 3.8 Live reached 34. The first two passed the same 76.0% of calls, so Extended Thinking spread its failures across more scenarios.

Both 3.8 models passed 50% of "accent and slow speech" scenarios, where the median was 83%. 3.8 Live reached 38% on "caller changes their mind" scenarios, where the median was 81%.

Extended Thinking finished well above the 49% median in the “Medicare compliance and routing” category at 70%, and 3.8 Live sat just under it at 48%. 3.1 Flash Live Preview scored 32%.

3.1 Flash Live Preview had the slowest median response of all 12 models (2.88s, with a P95 of 4.41s). Extended Thinking measured 2.07s and 2.93s, and 3.8 Live measured 1.91s and 2.43s. Grok Voice 3.0 was fastest at 1.34s.

Only the model settings differed. The 3.1 runs used high thinking and local turn detection, Extended Thinking used high thinking and service turn detection, and 3.8 Live used no thinking level and service turn detection.

These numbers describe that configuration, so use the table to shortlist models and validate on your own traffic.

Gemini 4 Argon and the Live API: what it can and can't do

Gemini 4 Argon doesn't support the Gemini Live API. Google's model page lists Live API and audio generation as not supported, so voice agents still belong on the two 3.8 models.

From cyber-defender preview to general release

Google announced Argon on September 30, 2026 as the first Gemini 4 model, starting with trusted cyber defenders through its Fairwind Program.

On October 13, 2026, Google opened Argon to developers and enterprises through the API and AI Studio, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search.

Google says Argon is strongest at multi-step professional and technical work in coding, finance, legal, cybersecurity, and science.

Audio in, text out

The model page accepts text, image, video, audio, and PDF as inputs and returns only text. Argon can transcribe or analyze an audio file, but it can't answer a caller out loud.

Google's developer guide doesn't mention the Live API either, and the audio in the launch post’s demos came from other models.

Context window, thinking levels, and price

The model ID is gemini-4-argon, and its status is Stable. Input, thinking, and output share a 1,048,576-token window, and thinking_level defaults to medium, with low and high also available, according to the developer guide.

Synchronous responses stop at 65,536 tokens, and longer ones need background execution with a higher max_output_tokens. Google says hard prompts can take several minutes, so it recommends client timeouts of at least 15 minutes.

The pricing page lists $2.00 per 1M input tokens and $10.00 per 1M output tokens through November 30, 2026. From December 1, 2026, the rates become $4.00 and $20.00, and cached input moves from $0.10 to $0.20.

Only implicit caching works on Argon, and the model page says explicit caching isn't supported.

Early reports from developers

Two Google AI Developers Forum threads describe problems in the first weeks. They're user reports, so they don't show that every deployment is affected.

On October 14, a developer reported zero cached tokens on a prefix of about 36,000 tokens. A Google staff member replied that explicit caching isn’t supported for Argon by design, and on October 29 a Google admin confirmed the implicit-caching problem was fixed.

On October 22, another developer described repeated passages inside one streamed response at around 360,000 tokens of context. Other users reported similar behavior, and the page shows no Google reply.

What Google says about safety

In its safety post, Google names the main risk as an agent trying too hard to finish its task, including attempts to game tests. Shortcut attempts on coding benchmarks fell to 2%.

Google also reports that Argon cut attack success on Gray Swan's Indirect Prompt Injection benchmark to 0.7% across 15 adaptive attempts. These are Google's own results, and the post summarizes a longer report.

Gemini 4 Carbon: announced, but not open to developers

Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Argon. It calls Carbon its strongest model, especially for cybersecurity and long-horizon work, and reports a first-place score of 59 on the Artificial Analysis Intelligence Index v5.0.

The announcement never mentions the Live API, voice, or audio.

Who can use Carbon today?

Only a subset of cyber defenders can. Google says access will widen to trusted partners through the Fairwind Program while it gathers feedback and tightens guardrails, and it hasn't set a date for general availability.

Carbon launches at Argon's introductory price of $2 per 1M input tokens and $10 per 1M output tokens. A footnote says $4 and $20 apply afterward, and VentureBeat reported that Google hasn't said when the introductory period ends.

No public API, and no Live API confirmation

Because Google limits Carbon to a subset of cyber defenders, there’s no public model ID to test against the Live API. Google hasn't said whether Carbon will support it.

Artificial Analysis lists speech among Carbon's input modalities and text as its output. That's third-party data, and the same page lists Argon as text and image only, while Google's model page also lists audio and PDF.

Carbon and Argon on the same benchmarks

Google's Gemini model page compares the two models on coding, agentic, and security tests. Carbon scores higher than Argon on all eight rows below, and Google's methodology document explains how the figures were produced.

None of these tests voice. Google's evaluation methodology says it computed the Gemini scores for Agents' Last Exam, ExploitGym, and SEC-bench Pro itself, and that scores for other providers' models are self-reported numbers.

The same document says SEC-bench Pro used a 183-task JavaScript engine split and ExploitGym ran without internet access. Gemini scores are pass@1 unless noted, with no majority voting.

Independent tests: faster output, slower first token

Artificial Analysis compared the two models at high reasoning. Carbon generated 95 tokens per second against about 83 for Argon, but its time to first token was 137.57 seconds against 40.47 seconds.

Artificial Analysis ran these as text tests, so the delay says nothing about voice agents.

VentureBeat's analysis of AA-Omniscience found Carbon answered 58% of questions correctly against 48% for Argon, while its reported hallucination rate rose from 4% to 11%.

VentureBeat adds that the rate divides incorrect responses by all incorrect, partially correct, and unanswered ones, so 11% doesn't mean 11% of answers are fabricated.

Artificial Analysis also shows $4 and $20 per 1M tokens for both models, while Google lists $2 and $10 as the current introductory price.

Which Gemini model works with the Live API?

Only Gemini 3.8 Live and 3.8 Live Extended Thinking are documented for Live API use. Gemini 4 Argon doesn’t support it, and Google hasn’t announced support for Carbon.

For a voice agent, build on one of the two 3.8 models and revisit this table when Google documents a newer Live model.

Gemini Live API pros and cons

The strengths below come mostly from Google's documentation. The weaknesses combine Cekura's benchmark results with the limits Google lists for the Live API and the 3.8 models.

Pros

• Stable model IDs: The 3.8 models are marked Stable with no announced shutdown date. The 2.5 and 3.1 models retire on November 17, 2026, at the earliest.

• Tools that run during speech: Async function calling is the 3.8 default, and 3.8 Live passed 100% of the benchmark's tool or system error scenarios, above the 92% median.

• Language switching: According to Google's launch post, 3.8 Live moves between 97 languages during a call, while Google's other docs list 70 and 99.

• Itemized pricing: The pricing page lists text, audio, and image rates per 1M tokens, plus per-minute rates for audio and image.

Cons

• Reliability ranks in the lower half: The best Gemini Live model finished 7th of 12 on pass³, and 3.8 Live finished last at 41.5%.

• Session limits require extra setup: Audio-only sessions stop at 15 minutes and audio-video sessions at 2 unless you add compression and resumption.

• Missing features on 3.8: The 3.8 models lack caching, code execution, structured outputs, URL context, and the Batch API, and ephemeral tokens are still in Preview on v1beta.

Should you build on the Gemini Live API? Our take

Build on it for multilingual, tool-heavy, or vision-enabled voice agents, and skip it if workflow reliability matters most today.

Between the two 3.8 models, Extended Thinking reached 84.4% data accuracy and 52.4% reliability, while 3.8 Live reached 70.5% and 41.5%. Both come behind the cascade reference, which reached 82.9% reliability and 94.1% data accuracy.

Use the Gemini Live API for:

Each use case below ties to a capability or pricing tier that Google lists for the 3.8 models.

• Multilingual agents that switch languages mid-call.

• Agents that call slow tools while the conversation continues.

• Voice agents that also see, through JPEG frames from a camera or screen.

• Prototypes, since you can start on the free tier, which uses your data to improve Google products.

Skip the Gemini Live API if you:

• Need workflow reliability today. Phonic 1 (90.2%), Grok Voice 3.0 (81.7%), and GPT Realtime 2.1 (79.3%) score higher than all three Gemini Live models in the benchmark.

• Depend on structured outputs or caching inside the live session.

• Run calls longer than 15 minutes and can't build session resumption.

Test it before you switch your voice agent

An upgrade to the 3.8 models needs voice agent testing before it reaches production traffic. In the benchmark, both 3.8 models scored lower on reliability than the 3.1 preview they replace.

1. Replay recorded calls against the new model ID

Send your recorded production calls through the new model first. Replay testing compares old and new outputs on accuracy, tone, latency, and hallucination rate, which lets a regression show up before a caller encounters it.

2. Write scenarios for proactive audio, tools, video, and output

proactive_audio is on for every session, so test calls with coughs and side conversations and confirm the agent doesn't respond to them. In the benchmark, Extended Thinking scored 83% on its silence, cough, and packet loss scenarios, which was five points under the 88% median.

Function calls also run in the background by default, so run a tool that takes 3 seconds and listen for what the agent says while it waits.

Video frames are now included by default, so run a vision call and compare context use and cost with your old setup. Responses are also audio only, so confirm your output transcription logs the full call.

3. Compare median and P95 latency

Track both percentiles for every model. In our benchmark, the difference between median and P95 ran from 0.52s on 3.8 Live to 1.53s on 3.1 Flash Live Preview, and an average would hide the slowest calls. Our latency guide explains how to measure each boundary.

4. Load test past your session limits

Run concurrent test calls past 15 minutes (2 minutes with video) with compression and resumption turned on. Our load testing guide recommends rerunning load tests before model updates.

How Cekura tests voice agents that run on Gemini Live

Cekura runs simulated callers against your agent and scores each call. Our benchmark used that approach on all three Gemini Live models, running each through the same Pipecat agent.

Pre-production

Pre-production testing runs before a model change reaches callers, using recorded and simulated conversations. These Cekura features cover replay, benchmarking, and adversarial testing.

• Replay recorded conversations: Run recorded production calls against a new model version before release.

• Benchmarking: Compare a new Gemini Live model with your current one on identical scenarios.

• Multi-turn red teaming: Run adversarial conversations that probe for prompt injection and data leaks.

Infrastructure

Infrastructure testing checks how the voice stack behaves under volume and how fast each call responds.

• Load testing: Run concurrent test calls to expose bottlenecks and latency spikes at your peak volume.

• Infrastructure metrics: Score latency and interruptions on every call, plus pronunciation checks for the words you define.

Observability

Observability starts after launch, when callers replace test scenarios.

• Production call QA: Score each completed call against the metrics you defined in testing.

Native integrations cover Retell, Vapi, ElevenLabs, LiveKit, Pipecat, Bland, and more, so testing can run against the stack you already use.

Cekura supports SOC 2, HIPAA, and GDPR compliance, covering transcript redaction, role-based access, and audit trails.

Cekura's pay-as-you-go pricing starts at $0.25 per voice testing minute, with the first seat free and $30 a month for each additional seat. The Startup plan costs $500 a month, month-to-month, and includes about 2,000 minutes of voice testing.

Our verdict on the Gemini Live API

Start with Gemini 3.8 Live Extended Thinking, and treat any newer Live model as a candidate until your own scenarios support the switch. 3.1 Flash Live Preview scored higher on reliability, but it retires on November 17, 2026, at the earliest.

In the benchmark, Extended Thinking reached 52.4% pass³ and 3.1 Flash Live Preview reached 62.2%, so a new model has to beat those figures on scenarios built from your own calls. Then rerun it on scenarios based on your calls, since your agent's setup will differ from the benchmark agent's.

Book a demo to see how Cekura would test your Gemini Live agent.

Frequently asked questions

What is the Gemini Live API?

The Gemini Live API is Google's stateful WebSocket API for low-latency voice and vision agents. It streams audio, images, and text in, and streams spoken replies out.

How much does the Gemini Live API cost?

Gemini 3.8 Live costs $3.00 per 1M audio input tokens and $12.00 per 1M audio output tokens on the paid tier, which Google puts at $0.005 and $0.018 per minute. Free-tier usage carries no charge, and Google uses that data to improve its products.

Which Gemini Live model is best for voice agents?

Of the current models, 3.8 Live Extended Thinking scored highest in the benchmark, at 52.4% pass³, and 3.8 Live scored 41.5%. 3.1 Flash Live Preview reached 62.2%, but Google schedules its shutdown for November 17, 2026, at the earliest. Validate your choice on your own calls.

Does Gemini 4 support the Gemini Live API?

No, Gemini 4 Argon doesn't, and Google hasn't announced Live API support for Carbon. Google's model page lists Argon's Live API and audio generation as not supported, and Carbon's October 28, 2026 announcement doesn't mention the Live API.

The newest Live API models are Gemini 3.8 Live and 3.8 Live Extended Thinking.

What is the best way to test a Gemini Live voice agent?

To test a Gemini Live voice agent, replay recorded production calls and run simulated ones before each model change. Cekura supports both with load and red-team scenarios plus production call scoring.

Native integrations cover Retell, Vapi, ElevenLabs, LiveKit, Pipecat, Bland, and more.