Gemini 4 Hydrogen: Google's fast model, open to everyone

On November 12, Google released Gemini 4 Hydrogen, the third model in the Gemini 4 family. Koray Kavukcuoglu (Google DeepMind, Chief AI Architect at Google) announced it as a model for everyday coding and agentic tasks. The same day, Hydrogen started rolling out to the Gemini API, AI Studio, the Gemini app and Google Search.

What is new here from an AI point of view?

Argon and Carbon are models for long, deep reasoning. Hydrogen is there to execute: write code, run tests and check changes, without keeping you waiting. Google presents it as the successor to Gemini 3.8 Flash: according to the announcement, it finishes tasks 2.5 times faster and makes fabricated claims more than five times less often.

On the Artificial Analysis Intelligence Index, Hydrogen scores 51 points at its default thinking level, and Argon scores 50. So the “fast” model has overtaken the flagship that launched on October 13, a month earlier. At the new, highest level, xhigh, it reaches 54 points (there are four levels: low, medium, high and xhigh). Carbon is still a long way ahead: when Carbon was announced in October, Google quoted 59 points.

An update while we are at it: Carbon, which only selected companies got in October, has been available since November 10 in the API for paying developers and is rolling out to Google AI Ultra subscribers.

How much does Gemini 4 Hydrogen cost?

$1.50 per million input tokens and $7.50 per million output tokens. There is one rate across the whole 1M-token context window, with no surcharge for long prompts. A response can also run to a million tokens.

For an “everyday” model, that is not cheap. Carbon costs $2 and $10 during its promotional period (the same as Argon), so Hydrogen comes out a quarter cheaper. The gap will widen after November 30, when Carbon’s price goes up to $4 and $20.

How do you try Hydrogen?

1. In the Gemini app, pick it from the model list. On a free account, Hydrogen only kicks in through automatic model selection, when Gemini decides it fits the question.
2. In Google Search, it works for all users, globally.
3. In the Gemini API and AI Studio, it is available on every account tier, under the name gemini-4-hydrogen.
4. In Google Antigravity, it requires a paid Google AI Pro or Ultra plan, and the same goes for Gemini in Google Sheets.

For a quick test without writing any code, Google has put demos in AI Studio, such as Geminisland: a shared voxel island where anyone can add their own structures with a single prompt and then walk around them.

What is Orchestrator?

Alongside the model, the Antigravity team showed Orchestrator. It is a new entry in the model list that switches between Carbon and Hydrogen within a single session. Carbon starts in read-only mode, maps the dependencies and lays out a plan. Hydrogen makes the changes and runs the tests, and when it gets stuck, it hands the task back to Carbon, which revises the plan.

An ordinary router picks a model once, after the first prompt, and sticks with it to the end. The Antigravity team’s reasoning is that in long AI agent tasks the difficulty shifts along the way and often only shows up mid-task.

According to Google, Orchestrator’s success rate is close to Carbon’s, while it uses 24% less quota on average and finishes 19% faster. The announcement does not say exactly how many tasks it solves. Google AI Ultra subscribers get Orchestrator with Carbon, and the paid AI Pro plan runs a version with Argon.

What should you watch out for?

First, the “class”. According to the announcement, Hydrogen leads “in its class” on HLE-Diamond (58.8%) and AutomationBench-AA, but Google does not name the models it was compared against. Specific rankings appear where the leaderboard is public: 5th on Agent Arena and 1st on Vals Tax Agent Bench (52.3%). The only model compared by name is its predecessor: on Terminal-Bench 4.1, Hydrogen scores 57% and 3.8 Flash scores 18.2%.

Second, the benchmark version next to Orchestrator. The Google blog describes the same 24% and 19% as a Terminal-Bench 4.0 result, and the Antigravity blog as a Terminal-Bench 4.1 result. One of the posts has a typo ;)

Third, the safety numbers, because they come from Google’s own tests. Hydrogen is said to do more than it was asked up to 81% less often, to score 96.7% on an internal agent honesty test and to be more than 2.3 times more resistant to indirect prompt injection attacks on the Gray Swan benchmark. The details are in the model card, and I explain why these numbers matter for agents in my guide to AI agent security.

FAQ

Is Gemini 4 Hydrogen free?

In Google Search, yes, for everyone. In the Gemini app, a free account only gets it through automatic model selection. In the API, you pay $1.50 per million input tokens and $7.50 per million output tokens.

How is Gemini 4 Hydrogen different from Carbon and Argon?

Carbon and Argon are for the hardest, longest tasks. Hydrogen is for fast execution and costs less, and in Orchestrator it works with them as a pair: the bigger model plans, Hydrogen writes and tests.

What context window does Gemini 4 Hydrogen have?

A million tokens of input and up to a million tokens in the response, at one price regardless of prompt length.