Google's Gemini 4 Hydrogen is a cheaper, faster model free users can try today

What Google released

As iFanr reported, Google has launched Gemini 4 Hydrogen, the newest model in its Gemini 4 family. Google positions its siblings, Carbon and Argon, for hard reasoning, long-range planning and demanding software engineering. Hydrogen is pitched at quick answers, lower cost and agent work: writing code fast, calling tools, and carrying out everyday tasks that take many steps in a row.

It is the third Gemini 4 model in six weeks. It also accepts text, images, audio and video, has a 1 million token context window (a token is a small chunk of text, and the window is how much material the model can hold at once), and can produce up to 1 million tokens of output. Google says that lets it work through complex PDFs, spreadsheets, charts and long videos.

The numbers, and who produced them

Most of the headline scores come from Artificial Analysis, an outside testing group. At the default High reasoning setting, Hydrogen scored 51 on its intelligence index, ahead of Argon's 50. A new top setting called xhigh lifts that to 54, just one point below Claude Sonnet 5.5 and GPT-6.1 Sol at their own highest settings.

Price is the other half of the story. Hydrogen costs $1.5 per million input tokens and $7.5 per million output tokens, and the rate stays the same across the full 1 million token window. Artificial Analysis estimates a task at xhigh costs about $3.16, or about $1.87 at High, for a gap of only 3 index points. Its xhigh cost is roughly 32% of Claude Sonnet 5.5's top setting, though still around 2.5 times GPT-6.1 Sol's.

Google's own figures say Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at the default setting. They also show a jump on Terminal-Bench 4.1 from 18.2% to 57%, and more than 80% fewer made-up facts. Artificial Analysis measured output of about 200 tokens per second.

Where it falls short

It does not win everywhere. On Terminal-Bench 4.0, a test of command-line coding work, Hydrogen scored 54%, behind GPT-6.1 Sol at 56% and Claude Sonnet 5.5 at 64%.

It looks stronger on made-up answers. On Artificial Analysis's AA-Omniscience test, Hydrogen's hallucination rate, meaning how often it states things that are not true, was 8%. Claude Sonnet 5.5 and GPT-6.1 Sol scored 43% and 52%. Those are single tests, so treat them as signals rather than a verdict.

Models that share the work

Alongside Hydrogen, Google added a feature called Orchestrator to Antigravity, its coding tool. Instead of one model doing everything from design to testing, Orchestrator switches between models within a single session. Carbon can draw up the overall plan, while several Hydrogen agents write code, run tests and check results in separate workspaces at the same time.

On Terminal-Bench 4.0, Google says this approach kept success close to what Carbon achieves alone while using 24% fewer credits and finishing 19% faster. That logic explains why a cheap, fast model matters: the expensive one is saved for the hard thinking.

Developers also posted demos of 3D scenes and interactive web pages built from short prompts. Google showed an online shop generated in one pass and Google Slides decks the model checks for overlapping text and fixes. These are showcases, not controlled tests.

Who gets it, and what changes in Search

Hydrogen is rolling out worldwide across the Gemini app, Google Search, Google AI Studio, the Gemini API and business services. Google AI Studio offers it on all tiers, and free Gemini app users can choose it in the model picker. In Antigravity it is limited to paying Google AI Pro and Ultra subscribers.

Google calls the Search change its largest model upgrade so far. Robby Stein, Google's vice president of search product, said Hydrogen better understands complicated queries, including ones with several conditions, typos or casual wording. Search also expands one hard question into several related searches at once, and answers can include generated images, comparison tables and custom interactive tools, not just links.

For the Gemini app, Google says Hydrogen follows complex instructions, tone and formatting rules more closely and can pull from Gmail and Google Drive to build a report.