Google's Gemini 4 Hydrogen replaces Flash as its fast model, with a dial for thinking
Google's new fast model lets teams choose how hard it thinks and pairs with a bigger model for planning, which could cut the cost of routine business AI.
The short version
- Hydrogen replaces Gemini 3.8 Flash and is live in Search, the Gemini app, AI Studio, the Gemini API and Antigravity.
- Google says it runs about 2.5 times faster than 3.8 Flash at a high reasoning setting.
- Four reasoning levels: low, medium, high and a new xhigh, plus a one million token context window.
- API pricing is $1.50 per million input tokens and $7.50 per million output tokens.
- Performance figures come from Google and are unconfirmed; one early review found weaker agent results.
Why it matters
If you run AI on repeated work such as support replies, data checks or back-office tasks, the biggest bill driver is using a powerful model for every step. Hydrogen is built around the opposite habit: match the model to the job. The catch is that the savings have to be measured on your own tasks, not taken from a launch announcement.
What Google launched
Google released Gemini 4 Hydrogen on November 12, 2026, as Times Square Chronicles reported. It takes over from Gemini 3.8 Flash, which had been the company's quick, lower-priced model. Google aims it at everyday coding and agent work, meaning tasks where an AI carries out several steps in a row, such as research, analyzing data and running tests.
The rollout is wide. Hydrogen is available in Google Search worldwide, in the Gemini app, in Google AI Studio and the Gemini API (the tools developers use to build on Gemini), and in Antigravity, Google's development platform.
On speed, Google says Hydrogen is about 2.5 times faster than Gemini 3.8 Flash when both are set to a high level of reasoning. That is the company's own claim, and the source says it has not been independently confirmed.
A dial for how much it thinks
The feature that matters most to a budget is the reasoning setting. Hydrogen has four levels: low, medium, high and a new one called xhigh. A higher level means the model spends more effort working through a problem before it answers.
That gives a team a way to trade quality against cost. A simple form submission does not need the same depth as a tangled customer complaint, and the dial lets you treat them differently without switching products.
The model also handles a context window of one million tokens. A token is a small chunk of text, roughly a word or part of one, and the context window is how much material the model can keep in view at once. That is enough for very large documents or long-running jobs.
For developers, the listed prices are $1.50 per million input tokens and $7.50 per million output tokens. Input is what you send in; output is what the model writes back.
Planner and worker, not one model for all
The most transferable idea comes from Antigravity. A new tool there, called Orchestrator, uses Google's larger Gemini 4 Carbon model to plan a job. Hydrogen then does the hands-on parts, such as making edits and running tests.
Think of an architect and a crew. The architect is costly and you want them on the hard decisions, while the crew handles the repeatable work. Any company building AI workflows can copy the pattern, whichever vendor it uses.
For small and mid-sized businesses, the source argues this could lower the running cost of AI receptionists, support agents and back-office automation. Not every call or form deserves the strongest model, so routing routine work to a cheaper one is where the savings would come from.
Reasons to hold back on the hype
Google reports a score of 51 for Hydrogen on the Artificial Analysis Intelligence Index, a third-party benchmark that rolls many tests into one number. That edges out Google's own Argon model, which is listed at 50. Again, these figures are Google's, and a one-point gap is small.
Two warnings appear in the coverage. At least one European outlet argues that the new model makes the Flash tier more expensive than it used to be, so a price comparison against your current setup is worth doing before any switch. And one early independent review found weaker results on agent-style tasks than the headline benchmarks would suggest.
Benchmarks measure general ability. They do not tell you how the model handles your invoices, your customers or your odd edge cases, and that is where a cheap model can become costly if it needs retries or human cleanup.
How to judge it for your own work
The sensible test is a small, repeatable one. Take real examples of a task you run often, run them through Hydrogen at different reasoning levels, and compare against what you use now.
- Track errors, response time and cost for each finished task, not just the price per token.
- Send routine work to the fast model and keep bigger models for hard reasoning.
- Compare an actual bill against your current setup before committing.
The broader shift is from one model for everything toward several models, each matched to a job. Teams that learn to mix them will probably spend less than teams that always reach for the biggest option, but only testing can show how much on your workload.
What to watch
- Do independent tests confirm Google's speed and benchmark claims?
- Does Hydrogen's pricing really make the Flash tier costlier in practice, as one European outlet argues?
- Do agent-style tasks improve beyond the weaker results seen in an early review?
Try it
Pick a task you repeat often, run the same set of real examples through Hydrogen and your current model, and divide total spend by the number of usable results.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Times Square Chronicles · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
