Gemini 3.8 Live Tops the Speech-to-Speech Leaderboard — and Costs 10x Less Than GPT-Live-1
Agentic AIGoogle DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 — and the two numbers that define the launch are $0.005 per minute (Google) versus $0.05 per minute (OpenAI’s GPT-Live-1). That’s a 10x pricing differential at the API layer, which matters enormously if you’re building voice agents that run thousands of hours of inference per month.
At those rates, one hour of audio costs approximately $1.38 with Google and over $3.00 with OpenAI. The cost difference isn’t incremental — it changes the unit economics of real-time voice applications in ways that determine whether some product categories are viable at all.
The Benchmark Position
Gemini 3.8 Live Extended Thinking debuted at the top of the Artificial Analysis Speech-to-Speech Leaderboard with an 82.6% score, ahead of OpenAI’s GPT-Live-1. The leaderboard measures voice agent performance across standardized speech-to-speech tasks — not just transcription accuracy, but agentic capabilities like background API calls, context retention, and response coherence.
Ranking first on a public third-party benchmark while pricing below the leader is an unusual market entry. It’s more commonly the other direction: aggressive pricing from a product chasing a quality leader. Gemini 3.8 Live is claiming both simultaneously.
Capabilities
Both Gemini 3.8 Live models support:
- Voice agents with background API calls — the model can initiate and wait on API responses mid-conversation without losing the audio thread
- Visual input processing alongside audio
- Simultaneous listening and speaking, rather than sequential turn-taking
- 97+ languages
Access is available through the Gemini API and Google AI Studio, with sample applications on GitHub.
The Tradeoff Google Acknowledges
The comparison isn’t clean. The Decoder’s analysis notes that “OpenAI’s model should still deliver more natural conversations” due to its full duplex capability, and that GPT-Live-1 “sounds better.” Google has prioritized benchmark performance and price-per-minute efficiency over the qualitative voice naturalness that OpenAI currently leads on.
For enterprise applications where the primary metric is task completion accuracy (customer service routing, structured data extraction, agentic workflows where the voice is a modality rather than the product), the benchmark result and pricing may matter more than vocal naturalness. For consumer-facing applications where the quality of the voice experience is part of the product value, OpenAI’s qualitative lead still has weight.
The Pattern
This is the second significant Agentic AI voice model development in a week — following Apple’s announcement that Siri would be rebuilt on Gemini foundation models, which was covered in this space on September 15. The underlying trend across both stories: the model provider layer is commoditizing faster than expected, and the differentiation is shifting to deployment cost, API ergonomics, and integration surface — not raw model capability.
The So What
The 10x pricing gap is the part that changes roadmaps. If you’re building voice agents today, the choice between Gemini 3.8 Live and GPT-Live-1 is no longer just about which model sounds better — it’s about whether your voice product’s unit economics close at $0.005/min or require $0.05/min to be viable. At scale, that’s the difference between a profitable product and one that bleeds margin on every conversation.
Google doesn’t need to match OpenAI’s voice quality to win the enterprise voice API market. It needs to be good enough — and meaningfully cheaper. Based on the benchmark position and the pricing, it may already be there for most use cases.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →