Google launches Gemini 4 Hydrogen for coding and AI agents

Google DeepMind, Google's AI research division, released Gemini 4 Hydrogen yesterday. It is a faster and cheaper model than Google's two most powerful models, Gemini 4 Argon and Gemini 4 Carbon. Those two are built to think for a long time about the hardest problems. Hydrogen is meant for everyday work: writing and improving code, building websites and running AI agents, AI programs that carry out a task in several steps on their own.

The new model succeeds Gemini 3.8 Flash, Google's previous fast model. According to Google, Hydrogen with its default setting finishes a task 2.5 times as fast and makes things up less than a fifth as often. On a difficult coding test it solved 57 per cent of the tasks, compared with 18.2 per cent for 3.8 Flash. On the general ranking of the research firm Artificial Analysis, Hydrogen scores 51 points according to Google, just above Argon with 50, while costing less.

Where it can be used

Anyone can use Hydrogen in Google Search. Google calls it the biggest model upgrade of its search engine so far. The search engine now understands questions with several parts better, searches for all the sub-topics at once and shows images and comparison tables more often. In the Gemini app, Hydrogen is in the list of models; people who use the free version get it automatically when the app decides Hydrogen is the right model for the question.

Developers can try out the model in Google AI Studio or build it into their own software through the Gemini API, the connection for programmers.

This costs 1.50 dollars per million tokens of input and 7.50 dollars per million tokens of output. Tokens are the small pieces of a few letters into which an AI model splits text. Hydrogen can read up to a million tokens at once, and the price stays the same however long the input is. Users choose how long the model thinks, in four settings from low to a new extra-high setting. According to Google, this lets them weigh up how much time and money a task is worth.

Two models working together

Paying subscribers to Google AI Pro and Ultra can also use Hydrogen in Antigravity, Google's coding platform. Antigravity is also getting a new feature called Orchestrator, which switches between Carbon and Hydrogen within a single job. Carbon can then divide the work, while a group of Hydrogen agents builds and checks the changes, each agent in its own sealed-off workspace. According to Google, this approach completed the tasks on a coding test almost as often as Carbon, while using on average 24 per cent less of the usage allowance and finishing 19 per cent faster.

Google also says Hydrogen is safer when it works as an agent. Compared with 3.8 Flash, it goes beyond what the task asks for up to 81 per cent less often, and it scored 96.7 per cent on Google's own test of how honest agents are. All figures in this article come from Google itself. Because the model is widely available straight away, independent testers can now measure them for themselves.

What does this mean for the road to the singularity?

AI that writes code is already normal, and there are plenty of strong models. What is new is that a model built for fast and economical work scores about as high, according to Google, as the top model that came out a month earlier. For competitors the choice is then simple: match the price, or offer something a cheap model cannot do. Every round of that race makes smart computing cheaper again.

The division of labour is more interesting. If one smart model makes the plan and a team of fast models does the work and checks it at the same time, building software starts to look like a building site with a site manager and a crew. Careful checking then becomes the most important job, because a crew that builds faster can also make mistakes faster. If that works, custom software becomes so cheap that a GP practice, a sports club or a small business can have programs made that fit exactly. And because the software used to make AI is also just software, such a team could in time help improve AI itself faster.

What is the technological singularity? Read the explanation →