Google introduced Gemini 4 Hydrogen, succeeding its earlier Flash-tier models and improving code generation and multimodal reasoning while staying fast.
This brief is compiled automatically by AI and may contain errors. Check the linked sources.
Model Releases & Open Source
Google Launches Gemini 4 Hydrogen for Fast Everyday Coding and AI Agent WorkflowsGGoogle DeepMind
In a nutshell: Google introduced Gemini 4 Hydrogen, succeeding its earlier Flash-tier models and improving code generation and multimodal reasoning while staying fast.
Key highlights: It completes tasks 2.5x faster than Gemini 3.8 Flash and scores 51 to 54 on the Artificial Analysis Intelligence Index, surpassing Argon from the same family.
Specs and pricing: Four adjustable thinking levels; API pricing is $1.50 per million input tokens and $7.50 per million output tokens.
Availability: Live across Google Search, the Gemini app, Google AI Studio, and developer tools including Cursor and Devin.
Thinking Machines Open-Sources Inkling 2 Small WeightsTThinking Machines
In a nutshell: Thinking Machines, led by Mira Murati, released the Inkling 2 Small weights under Apache 2.0, with a strong showing on web development leaderboards.
License: The weights are available for download on Hugging Face under the permissive Apache 2.0 license.
Benchmarks: Scored 1647 on Code Arena: WebDev, ranking first among US open models and 239 points above its predecessor Inkling-Small.
Anthropic Launches Claude Meetings: AI That Joins Your Video CallsAAnthropic
In a nutshell: Anthropic launched Claude Meetings in beta, letting Claude join online meetings as a visible participant and build documents and whiteboards while people talk.
Live collaboration: Claude creates docs, visual whiteboards and live notes during the meeting, and attendees can edit and correct them as they go.
After the call: Summaries, action items, the full transcript and everything Claude built are saved in a dedicated space.
Initial support: Google Meet first, rolling out in beta to Team and Enterprise plans.
Google Opens Pre-Orders for Its Intelligent Eyewear Lineup with Deep Gemini IntegrationGGoogle
In a nutshell: Google’s new smart glasses, built with eyewear brands and Samsung, are now available for pre-order, promising seamless hands-free AI assistance.
Design and hardware partners: Frames are designed by Warby Parker and Gentle Monster, with hardware made by Samsung.
AI features and launch: Wearers can ask Gemini by voice to recognize their surroundings, translate in real time and manage schedules; shipping is expected to begin December 10.
Figma Launches 3D Transforms in Open Beta with AI Agent SupportFFigma
In a nutshell: Figma rolled out 3D rotation and perspective effects to all Full-seat users and opened Figma MCP for coding agents.
Non-destructive 3D editing: Rotate and apply perspective to vectors, frames and text on the X, Y and Z axes while keeping the original layers editable.
AI ecosystem: The Figma Agent helps with motion design, and design specs can be exported to external coding assistants via the Model Context Protocol (MCP).
GitHub Publishes a Practical Guide to Agent Skills for a Standard Skills DirectoryGGitHub
In a nutshell: GitHub released guidance on Agent Skills, a standard format that lets custom tools and instructions be reused across different AI agent platforms.
Standardized design: SKILL.md files package repetitive engineering tasks and instructions into versioned, reusable skill assets.
Cross-platform: The goal is to end scattered, tool-specific instruction formats across coding agents—write once, use everywhere.
NVIDIA Shake-Up: Jensen Huang Resigns as President and CEO; CFO Colette Kress Named Interim CEONNVIDIA
In a nutshell: Jensen Huang, who led NVIDIA for 33 years, resigned as president, CEO and board member effective immediately; the CFO takes over on an interim basis while the board searches for a successor.
Leadership changes: CFO Colette Kress also serves as interim CEO, and director Harvey C. Jones was named the company’s first-ever Chair of the Board; the board is weighing internal and external candidates.
Business continuity: The board says operations, the product roadmap and customer commitments continue as normal, and the scheduled earnings call will go ahead.
Anthropic Formally Begins Its Initial Public Offering (IPO) ProcessAAnthropic
In a nutshell: Anthropic announced it has filed a registration statement with the US Securities and Exchange Commission, marking its move into the public markets.
Listing process: Anthropic launched an IPO investor portal and has submitted a preliminary prospectus to the SEC.
Underwriters: Goldman Sachs and Morgan Stanley are the lead underwriters for the offering.
SpaceX Plans More Than $74 Billion for Two Giant AI Supercomputing Campuses in the USRReuters
In a nutshell: According to media reports, SpaceX plans to invest more than $74 billion in large AI compute campuses in Arizona and Texas.
Arizona campus: About $50 billion on a roughly 747-acre site in Cochise County, with up to 6 GW of power across four phases and dry cooling to reduce water use.
Texas Starbase site: More than $24 billion for a standalone compute facility with its own natural gas power plant for stable, independent energy.
Interaction-as-a-Judge: Letting AI Agents Evaluate Apps by Actually Using ThemAArena.ai
In a nutshell: The Arena.ai team proposed a new evaluation paradigm in which computer-use agents interact with real software and use that experience as a reinforcement learning signal.
A new way to evaluate: Compared with static tests or text-based judging, interaction traces markedly improve detection of bugs and of real user-experience preferences.
RL from interaction alone: Without access to source code, fine-tuning a Qwen model with interaction rewards raised app-build success from 81.3% to 88.5% and cut redundant code by 35.7%.
Open datasets: Released two human-verified trajectory datasets, ArenaWebdevPreferences and ArenaWebdevCorrectness.
OpenAI Publishes Research on Fuzzy Robustness in High-Stakes DomainsOOpenAI Alignment Team
In a nutshell: OpenAI showed how automated red teaming and more inference compute can strengthen frontier models’ safety in cybersecurity and other high-risk dual-use domains.
Automated adversarial training: Attacker models get more test-time compute to search for jailbreaks, and attacks must succeed across repeated trials (the worst@n criterion) to yield reliable adversarial examples.
Stronger defenses: Training on this red-team data with reinforcement learning (RL) significantly raised flagship models’ resistance to extreme malicious steering.
Expo Launches Expo Simulators: Cloud Simulators That Close the Loop for Coding AgentsEExpo
In a nutshell: Expo introduced cloud simulators built for AI agents, so coding agents can run mobile apps themselves and produce test records for human review.
The pain point: Coding agents write code quickly but struggle to verify results; Expo Simulators gives them remote iOS and Android environments to tap through and check on their own.
Instant feedback: Every test run is recorded, with screenshots, video and a live clickable web link for the team, so pull requests come with real-device verification.
