Google and OpenAI take divergent paths on voice agent architecture

Voice agents face a fundamental latency challenge whenever they need to perform substantive work. Within a five-day window, Google and OpenAI each released competing solutions that reflect starkly different engineering philosophies.

Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking via the Gemini API and Google AI Studio, arriving five days after OpenAI's GPT-Live-1 launch. Both systems allow voice agents to continue speaking while background processing occurs, yet they accomplish this through fundamentally different mechanisms.

Reasoning inside the session

Google's Extended Thinking variant maintains speech generation, reasoning, and tool execution within a unified stateful session, even as external API calls remain in flight. When a function is configured as NON_BLOCKING, Gemini can sustain its speech output while awaiting tool responses, posing clarifying questions or providing status updates in the interim. Once results arrive, the model resumes from that point.

Developers can adjust reasoning effort to low, medium, or high on a per-request basis. The standard Gemini 3.8 Live variant omits the extended reasoning phase to minimize latency and token consumption.

The same model powers Gemini Live in the consumer application. Google characterizes that offering as its "end-user focused offering more closely comparable to ChatGPT and Claude," distinguishing it from the developer-facing models themselves.

Multimodal capabilities extend to the audio models. According to Google, "Visual understanding is excellent," and users can "converse with the model seamlessly about whatever you show it."

Coordinating two separate layers

OpenAI's GPT-Live-1 delegates the full-duplex conversation to the voice model while assigning reasoning and tool execution to a separate backend system—whether that is GPT-6 Astra, a lighter option like Luna, or a third-party alternative.

Isolating backend operations preserves voice-layer responsiveness, with OpenAI reporting turn-taking latency around 800 milliseconds. The tradeoff requires developers to manage coordination between the two layers themselves, routing context via sideband channels and determining conversation behavior while background tasks execute. This orchestration responsibility rests entirely within the application layer.

Stale work when a user interrupts

Both designs encounter the same problem when users interrupt or change direction mid-request, leaving background work executing that may no longer serve any purpose.

In Gemini's case, such work remains within the same session, though developers gain limited visibility into when tool calls actually terminate. OpenAI places cleanup responsibility on developers, who must cancel pending operations and prevent outdated results from re-entering the conversation.

Google asserts that Gemini performs better under the unpredictable conditions voice agents encounter beyond controlled demonstrations. The company states that Extended Thinking handles "background noise, heavy accents, and unexpected interruptions better than competing models."

Per-minute costs diverge sharply

Standard Gemini 3.8 Live follows Gemini Live API pricing: $0.005 per minute of audio input and $0.018 per minute of output. Extended Thinking introduces reasoning token charges, with additional fees for inputs such as live video and documents.

GPT-Live-1 charges $0.05 per voice minute for the front-end voice layer alone. Backend reasoning, function calls, and external agent operations all incur separate billing. Like GPT-6 Astra's tunable reasoning settings, developers can adjust cost per call, but voice agents that regularly invoke more capable reasoning models will experience rapidly escalating expenses.

OpenAI also incorporates DeepMind's SynthID watermark into generated audio.

Benchmark numbers, with caveats

Gemini 3.8 Live Extended Thinking achieved 82.6 on Artificial Analysis' Speech-to-Speech Quality Index, with task completion rates of 68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark. Google highlights these results directly, stating that Extended Thinking "holds the #1 spot on the Speech-to-Speech Quality Index and leads on complex task-completion benchmarks."

GPT-Live-1, paired with GPT-6 Astra at medium reasoning effort, scored 86.2% Pass@1 on Tau3's spoken customer-service evaluation spanning airline, retail, and telecom sectors. On Full Duplex Bench, it exceeded GPT-Realtime-2.1 by 30 percentage points.

Different tests, different stacks

These results do not represent direct head-to-head comparisons. Google and OpenAI employed different tests and configurations, and Google acknowledges that certain comparisons pit developer models against finished consumer products.

Claude Voice remains outside this developer conversation. Anthropic provides voice in its consumer applications but does not currently offer a real-time speech-to-speech API equivalent to Gemini Live or GPT-Live-1. Developers constructing voice agents around Claude must assemble more of the voice infrastructure independently.

Google consolidates speech, reasoning, and tool execution into one session, reducing middleware complexity while binding developers more tightly to its runtime. OpenAI demands greater orchestration but grants developers more control over the models and tools operating beneath the voice layer.

A Google spokesperson clarified the positioning: "Today's models are more centered on giving developers/enterprises tools to build voice agents. ChatGPT and Claude voice mode are full products rather than models, so the comparison is not apples-to-apples."