Realtime AI News
Google's new speech model Gemini 3.8 Live supports real-time reasoning
Google has launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice models it says can reason in near real time and handle speech and thought simultaneously. The Extended Thinking model claims a new high of 82.6 on the Artificial Analysis Speech to Speech Quality Index, and both are available through the Gemini API and Google AI Studio.
Google is trying to address the latency problem that has held back voice-based AI agents with the launch of Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The company describes them as its most advanced voice processing models to date, capable of near-real-time reasoning and simultaneous speech-and-thought processing.
The two models can also execute third-party software tool calls in the background, Google said in a blog post. The design goal is to run tool and API calls while maintaining a natural pace of conversation, so an agent can keep talking with a user while it works on a task that was just assigned to it, instead of forcing the user to wait through repeated interruptions.
On benchmarks, Gemini 3.8 Live Extended Thinking achieved a new high score of 82.6 on the Artificial Analysis Speech to Speech Quality Index, surpassing GPT-Live-1-Astra and Grok Voice Think Fast 2.0. Gemini 3.8 Live ranked second on the alternative Speech Agent Arena benchmark and first on ServiceNow's EVA-Bench, while the Extended Thinking model scored 68.6% on T-Voice.
Both models support automatic language detection and can switch languages mid-conversation. Google says they can understand and generate speech in 97 languages, support near real-time visual grounding, and use early verbal cues such as "let me check that" to acknowledge a user's prompt in a more lifelike way.
Gemini 3.8 Live is available now through the Gemini API and Google AI Studio, and as an enterprise private preview in Gemini Enterprise and Search Live. Gemini 3.8 Live Extended Thinking is available through the same channels plus Google Workspace via Docs, Gmail and Keep for subscribers, and in the Gemini Live applications. Developers can also integrate the models through partner platforms including Vercel, Agora, LiveKit, Pipecat, Fishjam and Vision Agents.
On pricing, Google said the standard Gemini 3.8 Live model costs $0.005 per minute for audio inputs and $0.018 per minute for outputs. The Extended Thinking model also charges for reasoning tokens and for additional inputs such as video and documents. Generated audio carries an invisible SynthID watermark that can be used to detect misinformation.
Voice has become the next front in the agent race, and Google's pitch is that reasoning and tool use no longer have to happen in sequence with the conversation. The details worth watching are how the promised latency and pricing hold up in real deployments, and how quickly third-party developers build on the new models.
Why it matters
Running reasoning, tool calls and conversation in parallel attacks the biggest experience gap in voice agents, and if the latency and per-minute pricing hold up in production, voice could become a decisive entry point in the next round of agent competition.
Nearby Updates
All09/16, 06:44
AI agent certification startup AIUC raises $40M to start auditing frontier models
AIUC, the startup formally known as Artificial Intelligence Underwriting Company, has raised $40 million to extend its agent certification and insurance work up to frontier AI models. Its AIUC-1 standard puts each agent through roughly 5,000 tailored risk and attack combinations and recertifies it every quarter, addressing what the company calls a security-review bottleneck rather than a capability gap.
09/16, 06:24
Salesforce launches Koa, its first CRM reasoning model, built on NVIDIA Nemotron 3 Super
Salesforce and NVIDIA used Dreamforce to unveil Koa, the company's first CRM reasoning model for Agentforce, post-trained from NVIDIA Nemotron 3 Super on nearly three decades of CRM data and run entirely inside Salesforce's own infrastructure. Koa is already powering an internal Slack agent, with customer pilots starting in October and general availability expected in U.S. regions in winter 2026.
09/16, 05:47
The AI data center boom is colliding with cities scarred by big industry
TechCrunch reported on September 15 that the national outcry against AI data center construction has spread to Philadelphia, where officials raised the possibility of building in a neighborhood already impacted by a now-defunct oil refinery. The story frames the AI infrastructure boom as a collision with communities that still carry the scars of earlier heavy industry.
09/16, 04:47
Spain's data watchdog publicises its first AI agent-linked data breach report
Spain's data protection watchdog has publicised the country's first data breach report tied to an AI agent, saying it received the first notification of a personal data breach allegedly carried out by an artificial intelligence agent. The case pushes autonomous agent behaviour into the regulatory frame, raising immediate questions about attribution, logging and oversight.