Google’s new live model can reason and call tools while keeping a conversation active, and Salesforce gives it a route into governed enterprise data. At Google’s published audio rates, the model layer can cost only cents per call; the economic contest moves to successful task completion.
Google’s Gemini 3.8 Live family attacks one of the most damaging failure modes in automated customer service: the silence that appears when a voice agent has to think or wait for a back-end system.
The standard Gemini 3.8 Live model is built for low-latency bidirectional sessions. The Extended Thinking variant can continue the interaction while planning and asynchronous tool calls run in the background, using conversational fillers and explicit “in progress” states instead of forcing a long silent pause. Google lists support for native audio and video, function calling, search grounding and a 131,072-token input context.
The raw price is small enough to change what enterprise buyers should measure. Google lists audio input at $3 per million tokens, with an equivalent of about $0.005 per minute, and audio output at $12 per million tokens, or about $0.018 per minute. On those published equivalents, five full minutes of audio in plus audio out would amount to roughly 11.5 cents of model-layer audio cost before text tokens, tools, grounding, telephony, CRM activity, other infrastructure and human escalation. Actual bills depend on how much audio is processed and generated, but the order of magnitude is the point.
A failed five-minute call that ends with a human transfer can be much more expensive than a successful seven-minute call.
Salesforce changes the distribution equation
Salesforce and Google Cloud expanded their partnership at Dreamforce on Tuesday, connecting Salesforce’s headless architecture with Gemini Enterprise and bringing Salesforce workloads to Google Cloud through Hyperforce. Salesforce says agents on the two platforms can reason and act on shared data without bespoke point-to-point integration.
For an enterprise already storing customer records, service history, sales activity and business logic in Salesforce, that can remove part of the integration tax that normally separates a capable model from a deployable agent. It also gives Google access to workflows where the application layer might otherwise hide the identity of the model provider.
The strategic pressure falls on independent voice-agent vendors that differentiate mainly on latency or orchestration. They can still win on domain tuning, telephony, reliability, evaluation and workflow depth. But “our speech loop is faster” becomes a weaker moat when Google can combine native multimodal infrastructure, asynchronous tool use and a major enterprise distribution channel.
The best procurement metric is no longer tokens per dollar or even dollars per minute. It is fully loaded cost per correctly completed task, measured after tool failures, context growth, compliance review and human fallback. The gap between a cheap conversation and a cheap resolution is where most of the economics now sit.
Gemini 3.8 Live makes the model component cheaper and the interaction less brittle. Salesforce makes the route to operating data shorter. Together, those changes raise the bar for anyone selling voice automation on model access alone.
Sources
- Google AI for Developers, Thinking in the Live API: https://ai.google.dev/gemini-api/docs/live-api/thinking
- Google AI for Developers, Gemini Developer API pricing: https://ai.google.dev/gemini-api/docs/pricing
- Salesforce, September 15, 2026, Salesforce and Google Cloud Unify Infrastructure and Agents for One Connected AI Stack: https://www.salesforce.com/news/stories/salesforce-google-cloud-unify-infrastructure-and-agents/