Google has released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, positioning itself at the forefront of conversational AI that can talk, think, and handle background tasks without interrupting the flow of conversation. The new models support 97 languages and enable asynchronous tool calls while speaking, marking a significant step forward in making AI voice agents practical for real-world deployment.

How Is Google's New Voice AI Performing Against Competitors?

Google's latest Gemini models are delivering measurable performance gains in live voice interactions. Gemini 3.8 Live Extended Thinking (High) debuted at the top of the speech-to-speech index with a score of 82.6, ahead of GPT-Live-1 Astra at 81.5, and also ranked first on the Tau Voice benchmark at 68.6 percent. The standard Live model offers a cheaper and faster alternative, though it performs less effectively on complex agentic voice tasks that require reasoning and decision-making.

Pricing remains competitive in Google's favor. The standard Gemini 3.8 Live model costs approximately $0.84 per hour of input audio, while the Extended Thinking High variant runs $3.50 per hour, positioning both below several competing live voice models. This pricing strategy reinforces Google's focus on making production-grade voice agents economically viable for developers and enterprises.

The developer rollout includes integration with major platforms and frameworks. Gemini 3.8 Live is available through the Gemini API and AI Studio, with partner support from LiveKit, Pipecat, LangChain, and Vercel, expanding accessibility for teams building voice-first applications.

Why Are Decision-Focused AI Models Suddenly Competing With General-Purpose LLMs?

While Google's voice models grabbed headlines, a smaller startup called TypeSafe launched a fundamentally different type of AI model that is reshaping how companies think about AI deployment. Jev, created by Diogo Almeida and the TypeSafe team, represents a new category of AI optimized for making decisions rather than generating text. The model claims to be 20 to 200 times faster and 40 to 400 times cheaper than small frontier large language models (LLMs), with output tokens priced at zero.

"After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I've spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev," stated Diogo Almeida, co-founder of TypeSafe.

Diogo Almeida, Co-founder at TypeSafe

The key innovation behind Jev is a training method called RLCD, or reinforcement learning with calibrated decisions. Unlike traditional autoregressive LLMs that generate text token by token, Jev is designed to make structured choices, classify information, route requests, and score options without the overhead of free-form text generation. This architectural difference makes it fundamentally unsuitable for tasks like creative writing or open-ended conversation, but exceptionally efficient for production systems where AI needs to make binary or multi-choice decisions.

The practical implications are significant. Engineers and researchers have identified Jev as a potential replacement for LLMs used as structured classifiers, judges, and routing policies in production systems. The most plausible use case involves compiling expensive LLM calls into many smaller, task-specific AI functions, similar to how developers use DSPy-style signatures and typed prediction abstractions.

What Are the Key Differences Between Decision Models and Traditional AI?

• Output Format: Jev requires predefined output formats and cannot produce free-form text, making it a constrained inference engine rather than a general language model.
• Speed and Cost: The model delivers 20 to 200 times faster inference and 40 to 400 times lower costs compared to small frontier LLMs, with output tokens priced at zero.
• Calibration and Reliability: Jev is trained to provide calibrated decisions with no hallucination, meaning it produces confidence scores that accurately reflect its certainty rather than generating plausible-sounding but false information.
• Parallel Sampling: Unlike autoregressive models that generate one token at a time, Jev can sample multiple decision paths in parallel, enabling faster batch processing.

The TypeSafe launch resonated strongly with the AI community, sitting atop Hacker News for an entire day despite competing announcements from Periodic Labs and Google's own Gemini 3.8 Live release. This suggests that developers and engineers recognize a genuine gap in the AI tooling landscape: the need for fast, cheap, reliable decision-making systems that don't require the full power of a general-purpose LLM.

How Does This Shift Change the AI Infrastructure Landscape?

The emergence of specialized decision models like Jev points to a broader trend in AI deployment. Rather than using a single large model for all tasks, production systems are increasingly adopting a layered approach where expensive, capable models handle complex reasoning while cheaper, faster models handle routine decisions and classifications. This mirrors how software engineering has evolved, with specialized tools and libraries replacing monolithic frameworks.

The implications extend beyond individual companies. As more organizations adopt decision-focused models, the bottlenecks in AI infrastructure will shift from raw inference speed to reinforcement learning rollout throughput, verifier compute, and weight synchronization across distributed systems. This suggests that the next generation of AI infrastructure investments may focus less on raw compute power and more on the systems needed to train and deploy specialized models at scale.

Google's Gemini 3.8 Live represents the current frontier of general-purpose conversational AI, delivering state-of-the-art performance on voice benchmarks and competitive pricing for production deployment. However, the simultaneous rise of decision-focused models like Jev indicates that the AI market is fragmenting into specialized solutions optimized for specific tasks rather than consolidating around a single dominant architecture. For enterprises and developers, this fragmentation creates both opportunity and complexity, requiring careful evaluation of which model type best fits each use case in their systems.