Lead story / launch

Google rolls out a speed-focused Gemini model across Search and its apps

Google began rolling out Gemini 4 Hydrogen globally on November 12, bringing its speed-focused model to Search, the Gemini app, developer tools and business products. In Search, Google says the upgrade improves handling of multipart questions, typos and specific requests, adds more visual responses and surfaces a broader range of relevant links.

Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default setting and cuts fabricated claims by more than fivefold. It also reports a 57% score versus Flash’s 18.2% on Terminal-Bench 4.1. Those performance results come from Google’s published benchmarks.

For developers, API pricing is $1.50 per million input tokens and $7.50 per million output tokens, with flat rates across the full one-million-token context window. Four reasoning settings—low, medium, high and xhigh—let teams balance response time, reasoning depth and cost; high is the default. Access includes AI Studio and Android Studio’s paid tier.

Google’s Antigravity coding tool also gains Orchestrator, which switches between models such as Carbon and Hydrogen within a session. In another configuration, Carbon coordinates while Hydrogen agents build and verify changes in parallel, in isolated workspaces. Hydrogen can also tackle coding and debugging alone, making it both a standalone option and a supporting model rather than a replacement for every Carbon assignment.

infrastructure

Anthropic says AI got one Claude workload running on AMD in three days

Anthropic says Mythos wrote or adapted nearly all the chip-specific code for one Claude model’s prompt processing on AMD MI355X GPUs. Its investor materials put the work at roughly three days, versus six to twelve months with traditional engineering; the figures are not independently verified. The result points to potentially lower switching costs, not proof that every Claude workload can move freely between chip suppliers.

research

OpenAI says training on AI-generated attacks helped models resist jailbreaks

OpenAI trained models on adversarial prompts generated by an AI attacker and reports stronger resistance to fixed and adaptive jailbreaks. The controlled study covered cybersecurity, biology and violent activity. In a separate comparison, GPT-6 Astra answered safely on 91.5% of fixed cybersecurity jailbreaks, versus 59% for GPT-5.6 Sol. Those figures measure model responses to evaluated prompts, not the full safeguards used in production.

business

Some Nvidia employees urge Jensen Huang’s return, WIRED reports

WIRED reports that an internal post drew more than 1,600 comments, many supporting Jensen Huang’s reinstatement. Employees also asked why he left before the board’s investigation was complete. The figure counts comments, not individual employees seeking his return, and is not a companywide vote. Finance chief Colette Kress is interim CEO as the board searches for a permanent successor.

Also worth knowing

•  Thinking Machines makes its Inkling 2 Small model available to download

•  Attackers used AI to access systems or data at 10 organizations, Unit 42 says

•  Trump urged Jensen Huang to fight for his Nvidia job, the New York Times reports

•  OpenAI explains how it limits more than 500 agents working on their own

•  Anthropic tops Newcomer and Bessemer’s first ranking of 100 AI companies

•  Google adds five artist-built tools to Flow for camera views and character design

•  Chinese AI firms seek overseas Nvidia capacity amid fears of U.S. curbs, Reuters reports

•  Anthropic launches a Claude plugin for quotes and factory troubleshooting

•  Trump backs Tim Cook for Nvidia CEO, but Cook’s interest is unconfirmed

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

The Global Economy’s Tiny Support

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.