Agentic AIThe briefing
AI-generated illustration
Google’s new voice models can keep a conversation moving during background tasks. The useful test is whether dialogue, tool outcomes and access requirements line up.
Google announced Gemini 3.8 Live and Live Extended Thinking on September 15. Google describes Live as its option for scalable voice interaction, while Extended Thinking adds reasoning during conversation. These are publisher descriptions, not Ashlr test results. The announcement gives teams a reason to examine voice workflows where a user needs to keep talking while software works.
Google’s developer announcement describes asynchronous function calling: an agent can continue streaming audio while a tool or API request runs. It also lists visual context, multilingual interaction and incremental content updates. Extended Thinking supports configurable thinking during background work. Those details distinguish the release from a service that only transcribes speech or reads a completed text response aloud.
Both models are rolling out through the Gemini API and AI Studio. Consumer destinations differ: Search Live uses 3.8 Live; Gemini Live uses Extended Thinking. In Workspace, Extended Thinking is available in Docs for Google AI Pro and Ultra subscribers, and in Gmail and Keep for Google AI subscribers. Enterprise access is in private preview; customer-experience access and Extended Thinking for Workspace business customers are listed as forthcoming.
The developer post also discusses Gemini 3.5 Transcribe, a separate speech-to-text model released earlier. A team seeking a transcript should distinguish that requirement from a conversational agent that reasons and calls tools. Google’s demonstrations and published benchmark claims do not establish task accuracy, end-to-end response time or operating cost for your own workflow. Ashlr has not independently evaluated these models.
Consider an illustrative appointment lookup: the caller asks for an available time, then changes the requested day while the lookup is pending. Test whether the application cancels, replaces or clearly supersedes the old request before presenting its result. Keep proposed, pending and confirmed actions distinct in the conversation and the underlying records. For this trial, a sensible acceptance rule would be that every spoken confirmation corresponds to an actual tool result, and a failed lookup never becomes an invented appointment. Include an interruption, a delayed response, an ambiguous date and a backend error. Compare both model variants on the same cases rather than assuming that more reasoning is always the better fit. These are suggested checks, not measured results or claims about a particular API’s cancellation guarantees.
The practical opportunity is a conversation that can remain useful during work, but a pleasant voice is only one part of the service. Ask which requests the assistant may complete, which require confirmation, and how a person takes over when the result is uncertain. Keep your product and subscription in the decision: a consumer feature demonstration does not establish enterprise availability. For a pilot, review completed tasks, incorrect confirmations, handoffs and total handling effort together. Account for the surrounding integration and review work instead of treating a model-level claim as the project’s return on investment. A polished exchange that leaves an operator correcting records has not demonstrated the intended improvement.
Choose one reversible, low-stakes voice workflow and write a short script with both an ordinary request and a mid-request change. Verify the model and access route available to your team. Run the same script against each eligible variant with test data, then inspect the transcript beside the actual tool records. Decide in advance which mismatch would stop the pilot. Expand only after you can explain both a successful interaction and how a failed one returns to a person.
Source material behind this briefing. Our practical implications are Ashlr’s interpretation.
Corrected availability: the two models have different consumer destinations and Workspace eligibility. Enterprise access is in private preview, with additional enterprise surfaces still forthcoming. Removed unattributed claims of cost efficiency.
This briefing was prepared with AI assistance from the linked sources. The analysis may include interpretation; consult the original sources for complete context. Cover artwork is illustrative.
Ashlr release
AI-generated illustration
Phantom gained agent-native fleet commands for lane targets, reassigns, and model sets. Multiple products received new documentation and regression tests.
Agentic AIThe briefing
AI-generated illustration
Genie 4 renders real-time simulated environments with physics and audio. Gemini models steer the scenes so agents can rehearse multi-step tasks before deployment. Access remains limited to trusted testers and planned Google AI Ultra subscribers.
Agentic AIThe briefing
AI-generated illustration
OpenAI launched GPT-6.1 Terra, a mid-tier model for computer use and visual reasoning tasks. It costs roughly half as much as GPT-5.6 Terra while running faster on agent workflows.
