Voice Agents Drop to About $1.38 an Hour: Google Ships Gemini 3.8 Live, Built on a Different Architecture From GPT-Live-1

Google DeepMind released two developer-facing audio models on September 15 — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — available through the Gemini API and AI Studio. They support more than 97 languages and can make API calls in the background, process visual input and keep talking at the same time, with Extended Thinking positioned to keep a conversation going while the model reasons through multistep tasks and runs tools in the background. Both are priced identically: $0.005 per minute of audio input and $0.018 per minute of output (derived from $3 per million input tokens and $12 per million output), working out to roughly $1.38 an hour — against OpenAI's GPT-Live-1 at $0.05 per minute, or at least $3 an hour. The architectural difference matters too: GPT-Live-1 is a full-duplex voice model that delegates reasoning and tool use to a backend model, billed separately, while Google's two are native speech-to-speech models. On Artificial Analysis' Speech-to-Speech Quality Index, Extended Thinking ranks first at 82.6. Boundaries to note: the models themselves are stable and generally available, but Google's broader Live API remains in preview, and the model card states they can hallucinate and may occasionally be slow or time out.

Half the Price Gap Comes From Architecture, Not Subsidy

$1.38 an hour against at least $3 reads easily as a price war. But the two are structurally different, and that is worth separating first. GPT-Live-1 is full duplex: it listens and speaks simultaneously, while the actual reasoning and tool calls are delegated to a backend frontier model — and that part is billed separately. So its "$0.05 per minute" is not the whole cost of a call; complex questions add the backend model's bill on top. Google's two are native speech-to-speech models, with reasoning inside the same model, where calling APIs in the background, processing visual input and continuing to talk are one capability. Which means cost estimates cannot stop at the per-minute rate: **price out one typical conversation end to end, with backend reasoning and tool calls included.** The more complex the dialogue, the further the two structures drift from what the rate card implies.

The Quality Evidence Points Two Ways Right Now

On Artificial Analysis' Speech-to-Speech Quality Index, Extended Thinking ranks first at 82.6, with component scores including 68.6% on τ-Voice, 35.1% on Sierra's τ-Voice-banking and 97.7% on Big Bench Audio. GPT-Live-1 sits at 69.8 on the same index, seventh of eleven models scored. There is another layer beyond the leaderboard. Analysis suggests GPT-Live-1 should still deliver more natural conversation thanks to full duplex — and that it sounds better in demos — concluding that Google optimized for price over quality. The two kinds of evidence do not conflict: one measures task completion and audio understanding, the other measures whether it sounds like a conversation. Teams building call-center QA or tool-driven voice agents care more about the first; anyone building companionship, sales, or anything that has to hold a person's attention cares more about the second. That 35.1% on τ-Voice-banking deserves its own note: it is the lowest figure in the set, and banking is one of the scenarios voice agents are most often expected to land in. Don't evaluate on the index total alone.

Two Boundaries to Confirm Before Production

First, **a stable model is not a stable pipeline.** These two models are generally available, but Google's broader Live API is still in preview. The realtime path you depend on is itself in a preview stage, which is a real constraint for any team that has to sign an SLA. Second, the failure modes Google states itself: the model card says both can hallucinate and may occasionally be slow or time out. For voice, a timeout is far worse than in text — in text a user waits, whereas three seconds of silence on a phone call reads as a dropped connection. The move is not to hope it doesn't happen but to design the fallback script and the escalate-to-human trigger first. One discrepancy to flag: at least one source notes Google has not published a dedicated rate card for 3.8 Live, and the per-minute figures above are derived from token pricing. If you are putting the cost into a budget, go by what your own bill actually charges.

via: TechRepublic, The Decoder, MarkTechPost, OpenAI: Introducing GPT-Live