Google has launched Gemini 3.8 Live and Live Extended Thinking models, designed to maintain useful conversations while agents perform tasks in real-time. The models combine visual context with background tool calls, reducing latency as a critical design factor.
TL;DR
- Google introduces Gemini 3.8 Live models for real-time agent interactions, combining visual context and tool calls.
- OpenAI's Astra for Law improves legal workflows with specialized search and tools, achieving 54% correctness on a validation set.
- Figure's Helix 2.5 demonstrates improved task success in unfamiliar homes, with pretraining on human-behavior data.
What happened
Google's new Gemini 3.8 Live and Live Extended Thinking models are designed to keep conversations alive while agents perform tasks. The models combine visual context with background tool calls, allowing agents to reason and speak concurrently. This reduces latency as a central design constraint, making voice a practical control surface for agents.
OpenAI has introduced Astra for Law, a GPT-6 Astra configuration tailored for legal work. It includes legal instructions, specialized search, and tools for legal workflows. On a private Legal Research Bench validation set, Astra for Law achieved 54% correctness, compared to 38.7% for Astra using ordinary web search.
Figure's Helix 2.5 policy demonstrates improved task success in unfamiliar homes. Pretraining on the Index human-behavior dataset raised complete-task success from 9% to 56%. The policy was tested across 30 unseen homes, showcasing the transfer of broad human experience to robotic training.
Why it matters
For developers, Google's Gemini 3.8 Live models offer a new way to design agents that can perform tasks in real-time, reducing latency and improving user experience. This could lead to more practical and efficient voice-controlled agents.
For legal professionals and startups in the legal tech space, OpenAI's Astra for Law provides a specialized tool that can improve the efficiency and accuracy of legal research. This could be a significant step forward in automating legal workflows.
For investors and robotics startups, Figure's Helix 2.5 demonstrates the potential of pretraining on human-behavior data to improve robotic performance in unfamiliar environments. This could open up new opportunities for robotic applications in various settings.
Key facts
- Google's Gemini 3.8 Live models combine visual context with background tool calls for real-time interactions.
- OpenAI's Astra for Law achieved 54% correctness on a private Legal Research Bench validation set.
- Figure's Helix 2.5 raised complete-task success from 9% to 56% with pretraining on the Index human-behavior dataset.
- Crusoe raised $3.9 billion in Series F funding at a $30.9 billion valuation to scale AI infrastructure.
- Profound raised $180 million in Series D funding at a $1.8 billion valuation for its AEO/marketing platform.
- Nvidia CEO Jensen Huang opposed antitrust waivers for AI labs, framing AI safety as an engineering problem.
- OpenAI is in early talks for a funding round that could value the company at more than $1.2 trillion.
- Emulate, a UK startup, is in talks to raise about $700 million in seed funding for physical-world behavior simulation.
Context
The recent developments in AI highlight the importance of connecting intelligence to its operating environment. Conversation requires timing, legal work requires authoritative context, and robotics requires transfer across messy physical settings. All three require computation that someone can actually deliver.
As models acquire more responsibility, progress will increasingly be measured by completed work under real constraints. The fascinating part is how much invention remains between an impressive model and a system we can comfortably depend on.
The AI landscape is rapidly evolving, with significant investments and advancements in various sectors. From real-time agent interactions to specialized legal tools and improved robotic performance, the opportunities extend across the entire chain of AI development and deployment.
