OpenAI’s voice model doesn’t think. That’s the point.

Google and OpenAI both recently unveiled voice models designed for real-time, interactive voice assistants. Within just five days of OpenAI releasing GPT-Live-1, Google responded with Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on the same day. Both systems allow a voice agent to keep talking while it performs background tasks, but they approach this functionality in very different ways.

Gemini 3.8 Live Extended Thinking keeps reasoning and tool execution within the voice model itself, enabling it to continue speaking while it asynchronously sends requests to external APIs. This design allows the model to maintain a single stateful session, handling speech, reasoning, and tool execution together even when external calls are ongoing.

Developers can choose the level of reasoning effort, from low to high, to balance latency and token cost. However, this approach means developers must manage the coordination between the voice model and the reasoning layer, passing context through sideband channels and deciding what to do while background tasks run.

In contrast, OpenAI separates the voice model from the reasoning engine. GPT-Live-1 handles the real-time conversation, while a backend model like GPT-6 Astra or a third-party option takes care of complex reasoning and tool execution separately. This design keeps the voice layer responsive, with turn-taking latency around 800 milliseconds, but it places the burden of orchestration on the application layer.

Developers must manage context passing, job cancellation, and prevent outdated responses from contaminating the conversation.

Google claims its Extended Thinking model handles real-world voice agent scenarios better than competing products. It demonstrates better performance with background noise, heavy accents, and unexpected interruptions. In terms of per-minute costs, the standard Gemini 3.8 Live model is priced at $0.005 for audio input and $0.018 for output.

Extended Thinking, however, adds additional charges for reasoning tokens and other inputs like video or documents. OpenAI's GPT-Live-1, on the other hand, costs $0.05 per voice minute for the front-end layer alone. The backend reasoning and external operations are billed separately.

Benchmark tests show Gemini 3.8 Live Extended Thinking excelling in speech-to-speech quality, scoring 82.6 on the Artificial Analysis index and achieving high completion rates on various benchmark tests. GPT-Live-1 paired with GPT-6 Astra scored slightly higher on some tests, but the different testing methodologies make direct comparisons challenging.

Google emphasizes that its new voice model leads in complex task completion benchmarks. However, these results come from different tests and setups, and the comparison is not a simple apples-to-apples matchup.