Google launches Gemini 3.8 Live to take on OpenAI's GPT-Live-1 at a fraction of the cost

What happened

Google DeepMind has released two new audio models for developers, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, positioning them directly against OpenAI's GPT-Live-1 in the voice AI market. Both new Gemini models currently top the Artificial Analysis speech-to-speech leaderboard, a widely referenced independent benchmark for conversational AI quality.

The headline differentiator is price: Google has set the cost at $1.38 per hour of voice conversation, substantially undercutting OpenAI's equivalent offering. According to reporting from The Decoder, GPT-Live-1 retains an edge in conversational naturalness because it supports full duplex audio, allowing both parties to speak and be heard simultaneously, a capability that typically makes exchanges feel closer to genuine human dialogue.

Why it matters

This is fundamentally a technology and platform story: Google is competing on benchmark performance and unit economics for the infrastructure layer that will power the next generation of voice assistants, customer service bots and embedded conversational agents. Cheaper, benchmark-leading speech-to-speech models lower the cost floor for any organisation building voice-based products, which should accelerate experimentation and deployment across sectors that have so far treated voice AI as too expensive or too immature to scale.

For technology and transformation leaders, the more interesting signal is the trade-off being made explicit: cost and raw benchmark performance versus conversational realism. Full duplex audio is not a cosmetic feature — it governs how natural turn-taking, interruption and overlap feel in a live exchange, which matters enormously wherever voice AI sits in a customer-facing or employee-facing role.

By the numbers

• $1.38 per hour of voice conversation is Google's price point for Gemini 3.8 Live, positioned well below OpenAI's GPT-Live-1.
• Two new models were released: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking.
• First place on the Artificial Analysis speech-to-speech leaderboard is currently held by the new Gemini models.

The Renascence take

Benchmark leadership and low cost will win headlines, but they are not the same thing as a good conversation. Voice interfaces live or die on the texture of interaction, not the price per hour or a leaderboard score.

Most organisations evaluating voice AI will default to comparing price sheets and benchmark rankings, because those are easy to put in a slide. The variable that actually determines whether a customer trusts and tolerates a voice agent is far subtler: can it be interrupted, does it pause naturally, does it recover gracefully when two people talk at once. Full duplex versus the cheaper, turn-based alternative is really a proxy for how much friction a brand is willing to impose on its customers in exchange for lower infrastructure cost. Any operator piloting these models should test them with real interruptions, cross-talk and hesitation built into the script, not clean, scripted demos — because that is where the cost savings either prove irrelevant or turn into a tangible experience gap.