Blog · 13 November 2026 · 7 min read

Is Gemini 4 Hydrogen safe? Model card, safeguards, your data

Gemini 4 Hydrogen launched on 12 November 2026 for everyone. What its model card says about safety, what is still missing, and what happens to your data.

For everyone. Report pending. Same data rules.

Short answer: for everyday writing, research and coding with a person checking the result, Gemini 4 Hydrogen is not riskier than the Gemini models you may already use, and on Google’s own tests it is better behaved than the model it replaces. Two things deserve attention. The full safety report is not out yet, only a model card. And Hydrogen is the first Gemini 4 model that reaches almost everyone, including free users and Google Search, so the question of what happens to what you type now applies to far more people.

Everything below comes from Google’s launch announcement and the Gemini 4 Hydrogen model card, both published on 12 November 2026, unless a sentence says otherwise. It was checked on 13 November.

What Gemini 4 Hydrogen is

Hydrogen is the fast, cheaper member of the Gemini 4 family. Carbon and Argon are the heavy models for long, hard reasoning; Hydrogen is meant for quick execution, and the model card says it is adapted from Carbon. It replaces Gemini 3.8 Flash. In the API it costs $1.50 per million input tokens and $7.50 per million output tokens, in US dollars, flat across a 1 million-token context window. It has four thinking levels: low, medium, high (the default) and a new xhigh.

Unlike Carbon and Argon, it did not start with a small group of security partners. It rolled out globally on launch day: in Google AI Studio and the Gemini API, in Gemini for business, in the Gemini app, in Google Antigravity for paid AI Pro and Ultra subscribers, and in Google Search for all users. A footnote adds that free Gemini app users get it through automatic model selection, when Google decides it suits the prompt. In Antigravity it can also work under Carbon as a fast worker model, the pattern AI agent orchestration explains.

What Google’s tests say

The launch post compares Hydrogen with Gemini 3.8 Flash, and the agent numbers are the ones that matter for anyone letting it act on their behalf. All of these are Google’s own figures:

  • Overreach in agentic tasks, doing more than was asked, down by up to 81%. Under-delivery down by up to 79%.
  • Misaligned actions and task shortcutting across tool-use and coding evaluations cut by 12 times.
  • When a task is impossible or blocked, up to 2 times less likely to bypass permission boundaries, and a 96.7% honesty rate on Google’s internal agent honesty test.
  • More than 2.3 times more resilient to indirect prompt injection, where hidden instructions in a page or file try to hijack the model, on the Gray Swan benchmark.
  • Fabricated claims cut by more than 5 times compared with 3.8 Flash.

The model card adds the content-safety results, and it is candid about one regression. Safe responses rose slightly for text (+0.80 percentage points) and multilingual prompts (+1.10), and clearly for images with text (+7.00). But unjustified refusals, where the model declines a harmless request, went up by 5.57 points, particularly for prompts with images. If Hydrogen refuses something reasonable, that is a known behaviour, not you doing something wrong.

What is still missing

The model card says the full Safety and Security Report for Hydrogen will be available in the coming days. Until then, the detail behind most of the numbers above is not public. On the frontier risks, the card says Hydrogen did not reach Google’s critical capability levels for biological or cyber misuse, but did reach the alert thresholds for both, and that mitigations stay in place. For the earlier models, those reports exist: is Gemini 4 Argon safe and is Gemini 4 Carbon safe read them in the same way.

The card also lists ordinary limits. Hydrogen can still hallucinate, its knowledge stops in March 2026, and it can be slow or time out at the higher thinking levels. None of that is unusual. It is a reminder that a fast model answering inside Google Search is still a model, and that how to check an AI answer when you are not the expert applies to it as much as to any other.

What happens to your data

Neither the announcement nor the model card has a data section. That is normal for Google: data handling follows the product and the plan you use, not the model name. The model card points developers to the Gemini API terms and business customers to the Google Cloud terms. If the line between a model and the app you open is fuzzy, AI agent vs LLM sorts it out.

In the Gemini app, the Gemini Apps privacy hub sets the rules. With Keep Activity on, chats auto-delete after 18 months by default, and you can change that to 3 or 36 months. With it off, and for temporary chats, Google keeps them with your account for 72 hours. Human reviewers read a subset of chats, and reviewed chats are kept for up to three years, even if you delete your activity. Google’s own advice is not to enter confidential information you would not want a reviewer to see.

Through the API, the Gemini API terms split on payment. On unpaid services, including Google AI Studio and the free quota, Google uses prompts and responses to improve its products, and human reviewers may read them. On paid services it does not, and logs them for a limited period only to detect misuse. In the European Economic Area, Switzerland and the UK, the paid-service data rules apply to everything, including AI Studio and the free quota. Google’s developer post covers the rest of the API changes.

One thing is new with Hydrogen: in Search, it now powers the AI answers for every user. Neither launch post says anything about how that data is handled, and neither mentions a change. Search falls under Google’s privacy policy and your account settings, as before. If you would not type something into a search box, a smarter model behind it is no reason to start.

What to do before you use it at work

  1. Check which account you are on. A personal Google account and a Gemini for business account fall under different terms, even when the model is the same.
  2. Look at Keep Activity in the Gemini app once, on purpose, and decide which way you want it.
  3. If you use the API at work, confirm the project is billed. The free tier may be used for training, unless you are in the EEA, Switzerland or the UK.
  4. Keep passwords, API keys and client data out of prompts and agent sessions. A higher honesty score does not change that.
  5. Ask for answers you can check. Write a prompt that works first try shows how to ask for sources and caveats up front.

For the practical side, how to use Gemini 4 Hydrogen covers where to find it in the app, Search and the API. For how another vendor answered the same questions this month, see is GPT-6.1 Terra safe. And for the model family Hydrogen belongs to, start with what Gemini 4 Argon is.

Coursium is an iOS app that teaches people to use AI at work: writing clear instructions, checking what comes back and applying it to everyday tasks. Model names change every few weeks; those habits carry over. If that is what you want to practise, have a look at Coursium.

Frequently asked questions

Is Gemini 4 Hydrogen safe to use for normal work?

For drafting, research and coding with a person reviewing the result, yes. Google’s tests show fewer fabricated claims and less agent overreach than Gemini 3.8 Flash. It can still hallucinate, and the full safety report has not been published yet, so the output needs checking.

Who can use Gemini 4 Hydrogen?

Since 12 November 2026, almost everyone. It is in Google Search for all users, in the Gemini app (free users get it through automatic model selection), in Google AI Studio and the Gemini API, in Gemini for business, and in Google Antigravity for paid AI Pro and Ultra subscribers.

Does Google train on what I type into Gemini 4 Hydrogen?

It depends on the product, not the model. In the Gemini app, a subset of chats is reviewed by humans and kept for up to three years. On unpaid Gemini API services, Google may use prompts to improve its products; on paid services it does not. In the EEA, Switzerland and the UK the paid-service rules apply to all API use.

Why does Gemini 4 Hydrogen refuse harmless requests?

Google’s model card reports that unjustified refusals rose by 5.57 percentage points compared with Gemini 3.8 Flash, especially for prompts with images. Rephrasing the request with more context is worth a try.