Gemini 4 Hydrogen ships, 2.5x faster with 1M-token window

Good morning. Jensen Huang's resignation has now pulled in a sitting cabinet secretary, a spy scandal, and what one outlet called an 'affair Singularity'. Meanwhile Google quietly shipped a new model and AWS let autonomous agents write 200 cloud service emulators from scratch. It is a strange Friday to be in AI.

Today's reading time is 4 minutes.

Listen · Snippt Daily0:00 / 3:54

INDUSTRY

Google's fastest publicly available model before today topped out at Gemini 3.8 Flash, which set the bar for speed in agentic workflows but had no thinking modes and a shorter context window.

Driving the news: Google released Gemini 4 Hydrogen yesterday, a fast agentic model reporting 2.5x the speed of Gemini 3.8 Flash. It carries a 1 million token context window, four selectable thinking levels, and scores 51 on Artificial Analysis and 57% on Terminal-Bench 4.1. Google is also rolling it out inside Search globally, replacing the previous AI Overviews layer with what it describes as more visual and intuitive responses.

Four thinking levels let developers trade latency for reasoning depth on a per-request basis.
The 1M-token window is large enough to process roughly 750,000 words, or a full codebase, in a single call.

Zoom in: Google has been shipping Gemini variants in quick succession since opening Gemini 4 Carbon to everyone earlier this week. Hydrogen appears to be a separate, speed-optimized branch rather than a replacement for Carbon, targeting developers who need fast inference over maximum reasoning depth. The Search integration is the more immediate consumer story: hundreds of millions of people will see Hydrogen responses before most developers have written a single API call against it.

Gemini 4 Carbon launched with agent swarm support; Hydrogen is positioned as the faster, lighter sibling.
Google DeepMind published the model announcement through its main blog, with a separate Search-specific post from Google's product team.

Why it matters: For anyone building agentic pipelines, a 2.5x speed gain on a model with four thinking modes means you can run more reasoning steps inside the same latency budget. The 1M-token window removes the chunking workarounds most teams use for long documents or large codebases. The Search rollout also means Hydrogen's citation behavior will be scrutinized at scale almost immediately.

Bottom line: Google shipped a faster model, put it in front of its entire Search audience on day one, and left every other fast model provider to explain the gap.

DeepMind ↗

Get this every weekday

Two minutes, 7am ET. No fluff.

Also happening

AWS released a local emulator of 200+ cloud services, and autonomous agents wrote all the code. AWS Emulator is a free, CC0-licensed tool that replicates AWS APIs locally without a cloud connection. It was built from 438 API models entirely by AI coding agents, not human engineers. At launch it passes 56.6% of 42,857 published fidelity tests, meaning it is accurate enough for development and testing but not a perfect replica. The CC0 license means anyone can use, modify, or redistribute it without restriction. (MarkTechPost)

AI chatbots directed shoppers to fake retail websites in a new study. A study by IP protection firm Podqi, shared with The Verge, prompted several major AI chatbots with shopping queries and found they recommended counterfeit or risky websites alongside legitimate ones. The study tested a handful of shopping scenarios and found the problem across multiple platforms. For anyone building shopping assistants or retail integrations on top of these models, the finding is a concrete liability: the models do not verify domain authenticity before citing a URL. (The Verge)

Hugging Face's CEO pledged to keep fighting for open source AI after Jensen Huang's exit. Nvidia's $12.9 billion acquisition of Hugging Face made Huang one of the most prominent corporate backers of open-weight models. With Huang gone, Hugging Face CEO Clem Delangue publicly committed to continuing that work regardless of what Nvidia's new leadership decides. The statement is partly reassurance for the open source community and partly a signal that Hugging Face intends to operate independently of whatever direction Nvidia takes next. Whether Nvidia's board shares Huang's enthusiasm for open weights is now an open question. (TechCrunch)

From the publisher

Snippt

Image, video and voice tools in one place

Open tool

On our radar

llama.cpp hit version 1.0.0 after years of b-prefixed builds, a milestone for the most widely used local inference runtime in the open source community.
Google DeepMind workers escalated their unionization effort to the UK government after direct talks with management stalled, according to Wired.
40% of game developers surveyed at GDC Festival of Gaming 2026 say they now use generative AI as part of their job, up from last year's figure, per The Verge.
Amplitude says its AI agent Wave, built on OpenAI models, has cut its product shipping cycle to 1 to 2 days by combining planning, coding, and impact measurement in one loop.
The White House is reportedly discussing whether Commerce Secretary Howard Lutnick should be pushed out over his handling of the Nvidia spy scandal surrounding Jensen Huang, per SCMP.
Qwen 3.8 Flash Next scored 29 on Artificial Analysis, placing it between two other frontier small models and closer to the top of the Qwen 3.8 family than the bottom.

Also happening

On our radar

AI worth acting on, in your inbox