A faster draft is useful only if it is the draft you can use. Google’s new Hydrogen model brings Gemini 4 to free Gemini users and promises quicker everyday work. The reporting so far points to three different questions: who gets it, what extra thinking costs, and how you judge the result.

Based on 8 reviewed sources · See sources

Free access comes through automatic selection

Google’s 12 November rollout announcement covers the Gemini app, Search, AI Studio and the Gemini API. It also puts Hydrogen in Antigravity, its agent workspace, for paid Google AI Pro and Ultra subscribers.

From the source · Google ·

Google’s original rollout post separates everyone’s access from paid-plan surfaces and developer routes. The app’s automatic-selection condition is explained below.

View the original post on X ↗

The app’s help page supplies the important detail for a free account: most prompts use Flash-Lite, while prompts requiring more reasoning may use Hydrogen when automatic selection is on. Switching that setting off means Flash-Lite handles the responses. Paid Google AI plans allow manual model selection; work and school availability depends on the Workspace licence.

Google describes a global rollout. Its web-app country list includes Belgium, France, Germany and the Netherlands; mobile availability can differ. Those conditions matter more than assuming every surface has the same access or controls.

The highest setting has a measurable price

For developers paying by use, Google lists $1.50 per million input tokens and $7.50 per million output tokens. Tokens are the pieces of text the service processes and produces. A rate card alone does not tell you the price of finishing a job: a model can produce more text or spend more time working through it.

Independent evaluator Artificial Analysis tests Hydrogen at different reasoning settings. Its Intelligence Index combines nine evaluations across knowledge work, coding, science and general knowledge. The suite is mainly English-language; a score is not a percentage of client briefs completed successfully.

At the highest setting, Hydrogen gains three index points over high, while average cost per test task rises from $1.87 to $3.16. That is a real trade-off in this evaluation, not a quote for your next report. The app’s help page lists low, medium and high controls, so the developer model’s extra-high setting should not be read as a free-app option.

The detail

What extra reasoning bought in one independent evaluation

Artificial Analysis launch snapshot, checked 13 November 2026. Different settings of the same model.

What extra reasoning bought in one independent evaluation. Artificial Analysis launch snapshot, checked 13 November 2026. Different settings of the same model.
Reasoning settingIntelligence Index scoreAverage cost per test task
Low27$0.29
Medium41$0.80
High51$1.87
Extra-high (xhigh)54$3.16
Index v5.0.1 scores and evaluation-task costs are different measures. These costs cover the evaluator’s task mix, not a Gemini subscription, a guaranteed customer workload cost or a measure of time saved.Source: Artificial Analysis’s original evaluation · 2026-11-13

A stopwatch cannot choose the result for you

Offloop, which describes its product as an agent workspace, posted a four-panel whale-scene comparison. It says the models received the same prompt and references. The team preferred Hydrogen’s movement, water and visual cohesion, while reporting that GPT-6.1 Sol was quicker and cheaper in that run.

The attached video makes its preference inspectable, but the full prompt and reasoning settings are not supplied in the post, and one panel is labelled as an Opus retry. This is a product developer’s own demonstration, not a controlled quality ranking. We have not reproduced it.

The distinction is useful for people commissioning work. A benchmark can expose a cost trade-off, and a demonstration can show why someone preferred an output. Neither decides whether a particular draft meets your brief. Hydrogen expands the choice of everyday tools; the result still needs a reader with a reason to keep it.

Sources

Written with AI assistance from the sources above. Analysis and practical implications are our interpretation. Our editorial approach.