Google released two new models focused on real-time speech-to-speech:
- Gemini 3.8 Live
- Gemini 3.8 Live Extended Thinking
Both immediately ranked #1 on the speech-to-speech index.
Why this is important:
Most voice AI still feels like a slightly smarter chatbot. These new models can reason while speaking and call tools mid-conversation. That combination is the missing piece for production-ready phone agents and in-app voice assistants.
Voice is no longer just a demo feature — it’s becoming a real product surface.
Other Major Model Releases
TypeSafe AI launched Jev
A non-generative “System One” decision model that outputs structured probabilities instead of long text responses. The company claims significantly lower cost and latency than frontier chat models.
This is a clear bet that many enterprise tasks are about decision-making, not content generation.
Salesforce announced Koa
A CRM reasoning model built with Nvidia for Agentforce, plus a deeper Claude integration that puts dozens of sales workflows directly inside Anthropic’s app.
The trend is clear: big software companies are no longer just using AI — they’re embedding it into the systems people already work in every day.
Research Highlights
- Shanghai AI Laboratory released the Atria Dawn Preview along with a ~744 billion parameter agentic mixture-of-experts model.
- New papers on interactive scientific agents (ScienceBuddy) and physics-grounded video generation (PhysStream) also dropped.
The research focus is shifting from pure question-answering toward agents that can experiment and world models that respect physical laws.
Big Funding Rounds This Week
| Company | Amount Raised | Valuation | Focus |
|---|---|---|---|
| Factory | — | $5 Billion | AI Coding Agents |
| Profound | $180 Million | $1.8 Billion | Enterprise Search |
| Anew Labs | $290 Million | $1.5 Billion | AI Drug Discovery |
| Hang Ten | $53 Million | — | AI Services |
Novo Nordisk also partnered with Anthropic to use Claude across drug discovery and development — a strong signal that pharma is treating frontier models as core R&D infrastructure.
The Safety Debate Is Heating Up
This may be the most important story of the week.
Several frontier lab CEOs and safety researchers are now openly calling to “pace” or slow the development of recursive self-improving systems. Reports of poorly contained agent behavior and high-profile resignations have added urgency.
OpenAI, Anthropic, and Google are coordinating on safety testing. OpenAI is also supporting outside evaluators in the proposed House FRONTIER Act.
But the pushback was immediate:
- House Speaker Mike Johnson rejected any moratorium, calling it a gift to China.
- Nvidia CEO Jensen Huang told regulators to stay out of safety decisions.
- Mark Zuckerberg argued that labs already have strong incentives to build safely.
- President Trump dismissed the warnings.
In Europe, Commission President Ursula von der Leyen said she would invite frontier labs to Brussels to discuss a potential slowdown.
The core tension is now clear:
Should the U.S. prioritize safety and pacing, or speed and competitiveness?
What This All Means
Two forces are colliding right now:
- Product capability is accelerating — especially in live voice and enterprise agents.
- Political and safety pressure is rising faster than at any previous point in the AI race.
Voice agents that can reason and use tools, coding agents raising at $5B valuations, and open-weight Chinese models at frontier scale are all happening at the same time as serious calls to slow down.
The next 12–18 months will likely be defined by how this tension gets resolved.
