Gemini 4 Hydrogen costs $1.50 per 1M input tokens, which is $2.14 CAD, and $7.50 per 1M output tokens, which is $10.69 CAD, priced flat across a 1M-token context window. Google released it on November 12, 2026, for everyday coding and agentic work. The price is the second most important number in the launch. The first is 51, the score Hydrogen posts at its default setting on the Artificial Analysis Intelligence Index, against 50 for Gemini 4 Argon, the heavier model above it in the same family. A fast model that outscores the frontier sibling at default settings changes what a buyer should test first, and the test is not the one the launch page leads with.
The direct answer: Gemini 4 Hydrogen is Google’s fast Gemini 4 model, live now in the Gemini API, AI Studio, Gemini for business, the Gemini app and Google Search. It costs $2.14 CAD per 1M input tokens and $10.69 CAD per 1M output tokens. It carries 4 thinking levels, and the level your team leaves it on will decide the real bill, because the top level scores higher and takes longer on every task.
- What Gemini 4 Hydrogen does at its default setting
- Why the thinking dial sets the real price
- What changes for a Canadian team this week
- The test to run before you move real work onto it
What Gemini 4 Hydrogen does at its default setting
At its default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, and 54 at its new xhigh level, against Argon’s 50. Google reports that it finishes tasks 2.5 times faster than Gemini 3.8 Flash at the high setting, and that its Terminal-Bench 4.1 score is 57%, against 18.2% for 3.8 Flash. Google also reports fabricated claims cut by more than 5x against the same older model. Each of those figures is the vendor’s own, published in the launch post and its model card, and a buyer should treat them as the claim to test rather than the result to bank.
The shape of the release matters as much as the scores. Hydrogen is not a preview. It sits in the model dropdown in Gemini for business, in the API, and in Google Search for all users, which means your customers in Canada already meet this model when they search, whether your firm has chosen it or not. Gemini 4 Carbon is out. The token price is the small number made the same point from the top of the family: the per-task cost decides, and the token price is the small print. Hydrogen moves that argument down the price ladder, to the model your team will call 1,000 times a day.
Why the thinking dial sets the real price
Hydrogen ships with 4 thinking levels: low, medium, high as the default, and the new xhigh. The levels trade time and tokens for depth. A classification job that runs at low costs a fraction of the same job left at xhigh, and most routine work never needs the top setting. Google’s own pairing tool makes the point from the other side. Its Orchestrator in Antigravity switches between Carbon and Hydrogen inside 1 session, and Google reports near-Carbon task success on Terminal-Bench 4.0 with 24% less quota used and work completed 19% faster on average. The saving comes from routing, not from a cheaper sticker.
This site has priced that routing lesson twice this month. Fable 5.2 cut its token price by 20%. Your bill may not fall showed a unit cut that effort settings can spend before it lands. Spotify’s plugin cut Claude Code token use by 90% in its own test. Your token bill is a settings problem showed the same shape inside a coding tool. Hydrogen adds a third version of the lesson: the dial is on the model itself now, and the default is high, 1 step below the most expensive setting. A team that never touches the dial has made a pricing decision by accident.
For high-volume judgement work, the family already has a cheaper door. Gemini Decision 1 answers 1,000 questions in 1 pass covers the Decisions API, built for batches of small calls. Hydrogen is the general tool beside it, priced for the work that needs a full model in the loop.
What changes for a Canadian team this week
The converted prices are the working numbers for a Canadian budget: $2.14 CAD per 1M input tokens and $10.69 CAD per 1M output tokens, at the Bank of Canada rate of November 12, 2026. Early partner testing on Google’s launch page reports work running 2x to 2.8x faster with token use down 12% to 23% against 3.8 Flash, on complex data and enterprise workflows. If those ranges hold on your own workload, a migration from 3.8 Flash pays for itself in time before it pays in tokens. If they do not hold, the dial is where the loss sits, and the dial is yours to set.
One caution rides with the speed. A model tuned for rapid execution sits inside agent loops, and Google reports its own guardrail figures for that work: overreach in agentic tasks down up to 81% and under-service down up to 79% against 3.8 Flash, in Google’s internal evaluations. Those are internal numbers, published by the vendor, and they describe the direction of the work rather than a result your firm can rely on. The reliance has to come from your own test, on your own tasks, with your own acceptance rule.
The test to run before you move real work onto it
Run 20 tasks your team already pays for, half at the default high level and half at xhigh, and score the outputs against the answers a person already approved. Record 3 numbers per task: the time, the token count, and whether the output passed without a redo. The model earns the migration when the cost per accepted output falls, not when the cost per 1M tokens falls. This lane ran no hands-on test of Hydrogen for this piece, and the launch figures above are Google’s. Your 20 tasks will tell you more than the launch page can.
The recap is short. The fast model now outscores the frontier model at default settings, the price converts to $2.14 CAD per 1M input tokens, and the thinking dial decides what your team actually spends. Speed is the headline. The dial is the decision.
Set the dial on 1 workflow this week, measure the cost per accepted output, and read Gemini 4 Carbon is out. The token price is the small number for the top-of-family version of the same test.
