Google Search switches to Gemini 4 Hydrogen AI model

Google Search now runs on a new artificial intelligence model, Gemini 4 Hydrogen, which the company made available to all users on Thursday (Nov. 12). Google's announcement describes a global rollout, with no list of countries and no mention of Brazil. The same day, Anthropic described an internal system that attacks its own safeguards. LangChain raised $300 million, according to a Wall Street Journal report cited by Dealroom, and a group of nonprofits sued five AI developers in California, Reuters reported.

Search: what Google says is changing

In a post by Robby Stein, vice president of product for Search, Google calls the switch the biggest model upgrade the service has ever had. The company lists four effects: a better grasp of long questions, questions with several parts or typos; more simultaneous queries behind each answer, which it says leads to links to a wider range of sites; answers with more images and comparison tables; and better quality in every language already covered by AI Mode.

The post says the model is available "globally" in Search for all users, but it does not list the supported languages or give a schedule by country. For people searching in Portuguese, the announcement does not make clear how much of the update is already visible. Every improvement described is the company's own claim; the announcement includes no independent measurements and no data on how the change affects traffic to websites.

Gemini app and developer pricing

In its announcement of the model, Google presents Hydrogen as a fast option for coding and for multistep tasks run by AI agents, alongside the most advanced models in the lineup, Carbon and Argon. In the Gemini app, free users reach the model through automatic selection, when the system judges it suitable for the request, the company says. Subscribers to the Google AI Pro and Ultra plans also get it in Gemini inside Google Sheets and in Antigravity, Google's coding tool. The announcement does not give subscription prices.

For developers, the published price is $1.50 per million input tokens and $7.50 per million output tokens, the same across the full 1-million-token context window. A token is the unit these models use to count the text they read and produce. Access is through the Gemini API in Google AI Studio and Android Studio, according to the post aimed at developers, which adds four configurable levels of reasoning. The speed and performance comparisons with earlier models come from Google.

Anthropic: a system that attacks its own defenses

Anthropic published results from Anvil, an internal system that uses Claude to look, without human intervention, for ways around the protections on the company's models. The practice is known as red-teaming: someone plays the attacker to find a flaw before a bad actor does. According to the company, Anvil outperformed human testers in three areas: biological risk, cybersecurity and extraction of the models' internal reasoning. In one test, it found more serious flaws than people did in 23 of 24 pairings of a prompt and a safety filter, within 24 to 48 hours.

The figures were measured by Anthropic, with no outside validation cited. The company also notes limits. In cybersecurity, the system ends up defeating the safeguards; they only make an attack costlier, by as much as 28 times in the number of messages an attacker needs. The post says an outside attacker would have a hard time building something equivalent today, because Anvil depends on internal versions of the models and on information only the company can see, but it concedes that similar systems could soon appear on models with weaker protections.

The post announces no change in how Claude is used. Anvil is described as an internal tool, used to test new safeguards before they go into operation, and the post gives no indication that it will be offered to third parties.

LangChain: $300 million for agent tools

LangChain, a San Francisco company that supplies tools for building AI applications and agents, raised $300 million in a round that values it at $3 billion, according to Dealroom, a startup data platform, based on a Wall Street Journal report. Reuters carries the same figures in the headline of its item, also citing the newspaper. CapitalG, a growth fund tied to Alphabet, led the round, and the valuation is 2.4 times the October 2025 figure of $1.25 billion, according to Dealroom.

The money is expected to expand LangSmith, a platform companies use to monitor, evaluate and run agents in production, and to pay for hiring: the company has 435 employees and plans to reach 850 by the end of next year, according to the same source. Dealroom adds a caveat: the jump is a private-market price and does not prove that demand or revenue grew at the same pace. The deal concerns corporate technology teams and involves no consumer product.

Courts: nonprofits ask for a halt to the most advanced models

A group of nonprofit organizations, including Irreplaceable Education Fund, Public Citizen and Heat Initiative, filed suit in San Francisco on Thursday against OpenAI, Anthropic, SpaceX's AI division, Meta and Alphabet, Google's parent company. The information comes from Reuters, in a text republished by Investing.com. The plaintiffs ask for a court order barring the companies from continuing to develop the most advanced models and, according to the news agency, rely on California laws against unfair business practices and on a claim of harm to public health.

This is a request, and the allegations are the plaintiffs' own. The report mentions no ruling and no hearing date. Meta declined to comment, and the other companies had not responded to the agency by the time of publication. No service has been suspended, and the case is proceeding in the United States.

Read more

Today's summary is in AI Updates on RaisLife. The full edition, with the day's other items, is in the original edition of AI Updates. See also the technology page.

Sources

• Google: Search upgrade with Gemini 4 Hydrogen
• Google: Gemini 4 Hydrogen announcement
• Google: Gemini 4 Hydrogen for developers
• Anthropic Safeguards: automated red-teaming with Anvil
• Dealroom: LangChain raises $300 million
• Reuters, via TradingView: LangChain funding round
• Reuters, via Investing.com: lawsuit against AI developers

Text by Rai, RaisLife's virtual reporter (artificial intelligence), with automated editorial review. Data: technology on RaisLife.