Google released Gemini 4 Hydrogen, calling it its biggest model improvement to Search yet, designed for fast execution across multi-step agentic workflows rather than deep reasoning. It completes tasks 2.5x faster than the prior fast model Gemini 3.8 Flash and even beats the stronger Gemini 4 Argon on one broad intelligence benchmark. The update is rolling out to all Search users globally and also lands in AI Studio, the Gemini API, the Gemini App and the Antigravity developer platform.
Agentic Workflow
智能体工作流
Execution paradigm where models autonomously plan, invoke tools and complete multi-step tasks.
Curated Intelligence
Google has fully opened its flagship Gemini 4 Carbon model to developers, enterprises and consumers less than two weeks after its debut. Carbon ranks first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, with task completion costs more than 2x lower than the next-best external model and pricing at a quarter of Fable 5.2. Google is shifting emphasis from knowledge QA toward real-world agentic task execution.
Australian corporate law firm Gilbert + Tobin moved from open-ended employee experimentation with AI to deploying ChatGPT Enterprise, ChatGPT Work and Codex, shifting from individual assistance to defined multistep agentic workflows with human review and approval. CEO Sam Nickless frames the goals as better client service, a better place to work, and freeing time for complex strategic work where expertise matters most.
Amap's marketing team shared an engineering practice that embeds AI into software quality assurance: a LangGraph-based Graph Engineering decision graph on the R&D side organizes multi-role code review and outputs risks plus verification conditions, while an AgentLoop workflow on the testing side handles change analysis, test case design and automation. Both capabilities are live separately, but a unified task protocol, cross-side handoff and delivery gates remain under construction.
Alibaba Cloud has launched Qovest, an AI investment research workbench for professional institutions, featuring a custom financial Agent framework that automates data retrieval and report generation. Its key differentiator is the 'Research Space' mechanism, which allows users to persist findings and set up continuous tracking of assets, eliminating repetitive manual work. The product is in early rollout, with an enterprise version coming soon for financial teams.
TypeSafe AI and Treg have open-sourced Jev, an AI agent designed for automated lead generation and contact verification. By decomposing natural language queries into structured criteria, retrieving data, and filtering irrelevant results, Jev reduces the cost per verified lead to $0.0089—approximately 85% cheaper than competitor Clay. The project has garnered over 6,400 GitHub stars, highlighting the practical utility of vertical-domain agents in precise data matching.
A developer used Claude Opus 5.5 and Sonnet to rapidly build HackerTube, a YouTube alternative frontend curated from Hacker News comments. The tool employs an automated workflow to clean nearly two decades of recommendation data, filtering 40k comments down to 6,454 high-quality videos to counter algorithmic slop. This project demonstrates the efficiency of AI agents in executing end-to-end tasks including data cleaning, application building, and deployment.
Genspark has officially launched Media Studio, a video agent workbench that orchestrates LLMs and multimodal models to automate short drama production. The tool supports one-click template replication and interactive storyboarding, significantly lowering the barrier for non-professional creators. A promotional discount on member credits is currently active.
Genspark officially launched Media Studio, a video agent workspace that orchestrates LLMs, image, and audio models to automate the workflow from script to final cut. The platform integrates over 70 model versions, supporting one-click replication of viral templates and multi-modal asset generation, significantly lowering the barrier to professional video production.
Google has released Gemini 4 Carbon, ranking first on the Artificial Analysis Intelligence Index v5.0. The model focuses on enhanced cybersecurity capabilities (vulnerability discovery/patching) and long-horizon agentic workflows, while significantly reducing over-execution rates and resisting prompt injection attacks. It is currently available via a restricted pilot program for security defenders.