Google says its AI business has crossed from pilots to operations: 75% of Cloud customers now use AI products, 330 customers processed more than 1 trillion tokens each in the past year, and model traffic exceeds 16 billion tokens per minute. The company used Cloud Next ’26 to turn that scale into a product pitch for Gemini Enterprise Agent Platform, a full runtime and governance layer for enterprise agents.
Why it matters: retrieval stacks are being pulled from text-only search into multimodal memory. Google AI Studio said Gemini Embedding 2 is generally available and covers text, image, video, audio, and documents through one model path.
Why it matters: Google is turning Vertex AI from a collection of services into a governed agent platform. The linked Google Cloud post says Model Garden gives access to more than 200 models, including Gemini 3.1 Pro, Lyria 3, Gemma 4, and Claude families.
Google has put Deep Research on Gemini 3.1 Pro, added MCP connections, and created a Max mode that searches more sources for harder research jobs. The April 21 preview targets finance and life sciences teams that need web evidence, uploaded files and licensed data in one workflow.
Google is moving Gemini image generation from prompt craft to account context. U.S. Google AI Plus, Pro and Ultra subscribers can opt in to use Google Photos and Nano Banana 2 for personalized images, with source visibility and reference controls built into the flow.
HN reacted because the failure mode is painfully familiar: an exposed or unrestricted client-side key, delayed cost reporting, and budget alerts that are not hard stops. The Google AI Developers Forum post says a Firebase AI Logic project saw €54,000+ in Gemini API usage within hours, pushing commenters into a broader argument about cloud billing safety for small teams.
Google is rolling out Skills in Gemini in Chrome so users can save prompts and rerun them on the current page or selected tabs. The feature starts on Mac, Windows, and ChromeOS for English-US desktop users, with confirmations before actions like adding calendar events or sending email.
HN focused less on the model drop and more on the hard robotics question: how fast does reasoning need to be before it is useful in the physical world? Google DeepMind frames Gemini Robotics-ER 1.6 around spatial reasoning, multi-view understanding, success detection, and instrument reading, while commenters zoomed in on gauge-reading demos, latency, and deployment reality.
Google’s new speech model moves control from hidden settings into the text itself: audio tags can steer style, pace, and delivery across 70+ languages. Gemini 3.1 Flash TTS is in preview through Gemini API, Google AI Studio, and Vertex AI, reaches Google Vids users, scores 1,211 Elo on Artificial Analysis, and watermarks outputs with SynthID.
Google has turned Gemini into a native Mac resident instead of a browser tab, with Option + Space bringing up the assistant and screen sharing extending it to whatever is open on the desktop. The April 15 launch also makes the app free globally for machines running macOS 15 or later.
Google DeepMind's latest robotics model pushes a hard industrial task from 23% to 93% accuracy when agentic vision is enabled, putting a concrete number on embodied reasoning progress. The April 14 release also puts Gemini Robotics-ER 1.6 into the Gemini API and Google AI Studio, so developers can test the upgrade immediately.
Google DeepMind is pushing embodied reasoning closer to deployable robotics, not just lab demos. In the linked thread and blog post, Gemini Robotics-ER 1.6 reaches 93% on instrument reading with agentic vision and improves injury-risk detection in video by 10% over Gemini 3.0 Flash.