Gemini 3.8 Live Extended Thinking vs GPT-Live-1 vs Grok Voice Think Fast 2.0: A Developer's Buying Guide

The field got crowded fast

Google DeepMind announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026. They are Google's newest native audio-to-audio dialogue models.

The launch came only five days after OpenAI added GPT-Live-1 to its API. xAI had already released Grok Voice Think Fast 2.0 on July 29.

Developers now have three new voice-agent options from the major AI labs within a few weeks.

All three are tackling the same basic problem. A lot of production voice agents still use a pipeline where speech-to-text handles the caller's audio, a language model decides what to say, and text-to-speech turns the response back into audio.

The pipeline works, but every handoff adds latency. Things that make a conversation feel natural, such as interruptions, "mm-hm" backchannels, or correcting yourself halfway through a sentence, can also get lost between stages.

The newer models move more of that work into the voice model. The important difference is where the actual reasoning happens.

Gemini handles reasoning inside the same voice model. GPT-Live-1 uses a separate backend model for reasoning and tool calls. Grok keeps reasoning inside its speech-to-speech model.

The difference affects more than the benchmark results. It changes the integration, the pricing model, and how predictable the bill is once you have real traffic.

Here is how the three models work, what the published benchmarks measure, what the pricing covers, and what those differences mean in a real voice-agent project.

What Google actually shipped

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are native audio-to-audio models. They can take audio, video, images, and text as input and return audio and text.

Google is positioning the two models for slightly different workloads.

Gemini 3.8 Live is the lower-cost, scale-oriented option. It combines conversational intelligence with real-time dialogue and visual grounding. It can process visual input in near real time and automatically switch between 97 languages during a conversation.

Gemini 3.8 Live Extended Thinking adds background reasoning for more complex tasks. The model can reason while it talks, use natural filler phrases such as "let me check that," and continue working on a multi-step task while keeping the conversation moving.

One of the most important developer details is in the model card.

Extended Thinking introduces an asynchronous reasoning protocol. That changes an assumption that many existing Live API integrations are likely to make.

Previously, turnComplete: true was a useful signal that the model had finished the current work and was waiting for the user. With Extended Thinking, that is no longer enough.

The server can send turnComplete while background reasoning is still running or while the model is waiting for an asynchronous tool call. Your client now needs to look at interaction_status, which can be IN_PROGRESS or IDLE.

Function calling has changed too. Extended Thinking only supports NON_BLOCKING execution, so a synchronous blocking function call returns an error.

The thinking_level setting can be low, medium, or high. There is no MINIMAL option on this model, and proactive_audio cannot be disabled.

Those changes matter during a migration. The model may look like an upgrade, but the session lifecycle is different enough that it needs proper testing.

Both models are rolling out to developers through the Gemini API and Google AI Studio. Google is also making them available in enterprise and consumer products, including some Workspace experiences for Google AI Pro and Ultra subscribers.

Google says every audio output carries a SynthID watermark.

There is also a growing ecosystem around the Live API. Google lists Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents as integration partners.

For an existing voice project, that can save a lot of engineering work because these tools handle parts of the WebRTC or WebSocket media layer for you. Google also names Salesforce, Genspark, and Lumeris among the early enterprise partners using the models in production trials.

Meet the competition: GPT-Live-1's benchmark is really two models

OpenAI added GPT-Live-1 to its API on September 10, 2026, after first offering the experience through ChatGPT.

It is a full-duplex model. In simple terms, it can listen and speak at the same time rather than waiting for a clean turn boundary. OpenAI says this design produced a 30-point improvement on its internal Full Duplex Bench compared with GPT-Realtime-2.1.

There is an important detail behind the benchmark numbers, though.

GPT-Live-1 is primarily the voice layer. It does not do the main reasoning by itself. Reasoning and tool calls are delegated to a separate backend model that you choose, such as GPT-6 Astra, Luna, Terra, or a third-party model through client-side delegation.

So when a benchmark says something like "GPT-Live-1 Astra," the result belongs to the combination of GPT-Live-1 and Astra. It is not a measurement of the voice layer alone.

The same split appears in the pricing.

The GPT-Live-1 voice layer costs $0.05 per minute, billed per second. The backend model is charged separately according to its own token pricing.

There are a few other limitations to account for. GPT-Live-1 has no free tier, runs through the dedicated v1/live/sessions endpoint rather than Realtime or Responses, and does not support structured outputs, image input, or video input.

Migration from the older Realtime API is different as well. Prompts are now split between session instructions and delegation instructions, tools use a new delegation schema, and manual turn control has been removed.

Meet the competition: Grok Voice Think Fast 2.0

xAI shipped Grok Voice Think Fast 2.0 on July 29, 2026. The grok-voice-latest alias moved to that version on August 5.

Its architecture is closer to Gemini than GPT-Live-1. It is a single speech-to-speech model, so you do not have to configure a separate backend model for reasoning.

Grok Voice Think Fast 2.0 supports more than 25 languages, real-time tool calling, and live web search. xAI reports that it uses about 60% fewer reasoning tokens per turn than the previous version.

xAI also reports a drop in time-to-first-audio from 1.25 seconds to 0.70 seconds.

Pricing is straightforward. Grok Voice Think Fast 2.0 costs $0.08 per minute of audio, covering both the user's audio and the model's response. Text input is charged separately at $0.004.

The previous Think Fast 1.0 remains listed at $0.05 per minute for projects that explicitly pin that version.

One deployment detail may matter before price. Speech-to-speech access currently runs from us-east-1, so teams with strict latency or data-residency requirements should check that against their deployment needs.

What the benchmarks actually say

Before comparing the scores, it helps to know what each benchmark measures.

The Speech to Speech Quality Index from Artificial Analysis combines several dimensions, including speech reasoning, agentic performance, arena preference, and task success. It is therefore broader than a simple speech-quality test.

Big Bench Audio focuses on speech reasoning. It is closer to a spoken version of a general knowledge and reasoning benchmark than a full customer-service simulation.

tau-Voice measures whether a voice agent successfully completes a defined task. Sierra's tau³-Banking applies a similar idea to banking-style customer-service tasks.

None of these benchmarks is simply asking which model sounds the most human. Some arena components capture human preference, but reasoning and task completion are also part of the picture.

Google's launch materials, along with independent coverage from MarkTechPost and OfficeChai, report a Speech to Speech Quality Index score of 82.6 for Gemini 3.8 Live Extended Thinking.

The same comparison reports 81.5 for GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort, and 81.3 for Grok Voice Think Fast 2.0 at its high setting.

Gemini 3.8 Live Extended Thinking is also reported at 97.7% on Big Bench Audio. The base Gemini 3.8 Live placed second in the Speech Agent Arena in the same launch coverage.

On tau-Voice, the reported figures are 68.6% for Gemini 3.8 Live Extended Thinking, 67.9% for GPT-Live-1 with Astra, and 56.5% for Grok.

Sierra's tau³-Banking comparison reports 35.1% for Gemini and 32.0% for GPT-Live-1 with Astra. The same source lists an xAI realtime entry at 16.5%.

The xAI figure needs a caveat. The comparison does not clearly say whether the entry is Think Fast 2.0 or an earlier xAI voice model, so that particular number should be treated cautiously.

The test setup matters too. These are vendor-published comparisons using specific configurations. GPT-Live-1 is paired with Astra at medium reasoning effort, while Grok is tested at high effort.

Change the backend or reasoning setting and the result can change too.

For current independent measurements, Artificial Analysis' live leaderboard is more useful than a single launch-day comparison. These models are still new, and outside testing is continuing.

The pricing architecture nobody puts in one table

The biggest pricing difference is not the headline number. It is what that number actually covers.

flowchart LR
    subgraph U["Unified speech-to-speech"]
        direction LR
        A["Microphone audio"] --> B["Single model:<br/>listens, reasons, speaks"]
        B --> C["Speaker audio"]
    end

    subgraph S["Split voice layer + backend"]
        direction LR
        D["Microphone audio"] --> E["GPT-Live-1<br/>voice layer only"]
        E <--> F["Backend model<br/>Astra / Luna / Terra"]
        E --> G["Speaker audio"]
    end

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking use the same pricing table on the Gemini API pricing page.

Google lists audio input at $3.00 per million tokens, which is roughly $0.005 per minute. Audio output is $12.00 per million tokens, or roughly $0.018 per minute.

Google also does not list a separate higher audio price for Extended Thinking. Turning on deeper background reasoning does not change the listed audio rate.

Grok Voice Think Fast 2.0 uses a flat $0.08 per minute of audio. Artificial Analysis reports that same rate as $4.80 per hour of input audio.

GPT-Live-1 starts at $0.05 per minute for the voice layer, but that is not the whole cost. The backend model is billed separately.

TokenCost's analysis estimates a typical fully loaded cost of about $3.05 per hour when GPT-Live-1 uses the cheaper Luna backend. Using Astra raises the estimate to about $5.16 per hour.

A usage example makes this easier to understand.

Imagine a 60-minute connected voice session where the microphone stays open the whole time. The model speaks for 10 to 30 minutes of that hour, which is a reasonable range for a support conversation.

Using the published Gemini audio rates, that works out to roughly $0.48 to $0.84 for the audio portion of the session.

That is only an estimate based on the stated assumptions, not a vendor quote. Actual billing depends on tokenization, audio duration, and how the session behaves.

What matters is how those charges are structured.

Gemini charges for the voice model that also handles the reasoning. GPT-Live-1 has a separate voice charge and backend charge, so your total depends on the reasoning model you choose. Grok has a single-model setup, but its flat audio price is higher than the Gemini example above.

Feature comparison at a glance

What the async protocol means for your code

If you are integrating Gemini Extended Thinking specifically, the biggest change is the session lifecycle.

You can no longer assume that turnComplete: true means the model has completely finished. The client needs to keep track of interaction_status and wait for the session to reach IDLE.

Function calls need the same approach because the model expects asynchronous, non-blocking execution.

A basic handler looks like this:

// Old assumption: turnComplete meant the model had gone idle.
// New reality: keep listening until interaction_status says otherwise.
socket.on('serverContent', (message) => {
  if (message.turnComplete) {
    playQueuedAudio(message.audio);
  }

  if (message.interactionStatus === 'IN_PROGRESS') {
    // Background reasoning or an async tool call is still running.
    // Do not close the mic or show an idle state yet.
    return;
  }

  if (message.interactionStatus === 'IDLE') {
    // The session is genuinely done and waiting on the user.
    setUiState('listening');
  }
});

The check is small, but the behavior matters. A client that closes the microphone or switches to an idle state as soon as turnComplete arrives could interrupt background work.

Before adding your own handling, check whether your existing integration layer already supports the new message fields. Google lists LiveKit, Pipecat, Agora, and Vercel among the Live API partners.

For the base Gemini 3.8 Live model, the migration is simpler. It follows the older turn-based pattern, so applications that do not need background reasoning have less session-lifecycle work to do.

Which one should you actually use?

There is no single configuration that fits every voice agent. A more useful comparison starts with what your application needs.

Cost-sensitive, high-volume voice agents. Think order status, appointment scheduling, simple support questions, or FAQ calls where the number of sessions matters more than complex reasoning. Gemini 3.8 Live is designed for this type of workload. Its listed pricing does not add a separate reasoning charge, and automatic switching across 97 languages is useful for applications serving multiple markets.

Complex, multi-step workflows. For booking across several systems, technical troubleshooting, or other tasks where the agent needs to call multiple tools before answering, Gemini 3.8 Live Extended Thinking is built around that kind of interaction. The tradeoff is the asynchronous session lifecycle, so the integration needs more careful handling than a basic Live API setup.

Teams that want control over the reasoning backend. GPT-Live-1's split architecture can make sense when you want to use a cheaper backend for simple requests and a stronger model for harder ones. The voice layer stays the same while you change the model doing the reasoning. The main thing to remember is that the $0.05 per minute price is only for the voice layer.

Teams using the xAI stack or needing live web search in the voice workflow. Grok Voice Think Fast 2.0 includes live web search and real-time tool calling in the same speech-to-speech model. Its reported time-to-first-audio is 0.70 seconds. The tradeoffs are its $0.08 per minute audio price and the current us-east-1 speech-to-speech deployment.

The bottom line

The recent voice-model launches show why benchmark tables need context.

Two models can post similar scores while using very different architectures. One may include the reasoning model in the same system, while another may rely on a separate backend. That difference affects both engineering and cost.

For Gemini, the interesting part is the combination of native audio and in-model reasoning without a separately listed audio premium for Extended Thinking.

GPT-Live-1 takes the opposite approach. It separates voice from reasoning, which gives developers more control over the backend but also makes the total cost harder to describe with one headline number.

Grok keeps the architecture unified and adds live web search, but its published flat audio rate is higher than the Gemini example used above.

For an actual project, do not rely on the leaderboard alone. Run the same workload through the models and measure the things that matter to your application: latency, interruption handling, tool-call behavior, task completion, and total cost.

These models are moving quickly. The numbers in this article are a September 2026 snapshot, so check the current vendor pricing and independent benchmarks before making a long-term commitment.