AI News

Google DeepMind’s new Gemini 3.8 Live voice models for developers beat OpenAI’s GPT-Live-1 on benchmarks while costing a fraction of the price.

OpenAI is urging developers to slim down their AI agent instructions for GPT-6 Astra. Overly long skill descriptions and legacy rules are holding the new model back.

OpenAI’s GPT-Live-1 is the first voice model that can listen and speak at once. Yelp and Speak report noticeably more natural conversations with customers.

OpenAI’s new Agents API offers sessions, tools, and sandbox connections as a service. That changes what teams still need to build for their own AI agents.

ChatGPT Images 2.5 adds more precise editing, sketches, and faster generation. The real question is control in a creative workflow, not the demo.

GPT-6 Astra completed Portal autonomously. The open test setup shows progress in AI agents, but also why it is not proof of general autonomy.

Claude is now available in Apple CarPlay. What the voice AI can do in the car, where its limits are, and why Siri remains essential.

OpenAI ships a prompting guide for GPT-6 Astra with a blocklist for filler phrases. The advice carries over to other models, and it exposes a deeper problem.

OpenAI’s new model blocks direct manipulation almost completely. But for attacks hidden inside documents, the attacker success rate still sits at 8.5 percent.

The AI week: OpenAI declares the AGI era, Google drives down model prices, three chatbots fail at once, and agents hack their way through Hugging Face.