Google DeepMind unveiled Gemini 4 Hydrogen on November 12, 2026, with a bill that is easy to remember: $1.50 per million input tokens and $7.50 per million output tokens, the same price at any point in a one-million-token context window. The Gemini 4 Hydrogen announcement is signed by Koray Kavukcuoglu, senior vice president of Google DeepMind, and places it below Gemini 4 Carbon and Gemini 4 Argon, the manufacturer's two frontier models for long-horizon reasoning, with the promise of fast execution, coding and agentic tasks.
The novelty is not only technical. Hydrogen drops the Flash surname the company used in the previous generation, and that name change sums up the strategy: instead of selling a cheap model and an expensive one, Google positions an intermediate model at a fixed price across the entire window and lets the user decide how much thinking they want. The four configurable thinking levels are the lever, and the price no longer moves with them.
For whoever pays the API bill, the relevant question is not whether Hydrogen beats Argon on a benchmark, but whether finishing a task costs less. That is where the flat price changes the math.
An intermediate model with a new name
The hierarchy described in the announcement has three rungs. Carbon and Argon sit at the top, reserved for hard problems and long-horizon reasoning, while Hydrogen is presented as the execution engine for day-to-day work: writing code, running tests, verifying changes and resolving issues in repositories. The accessibility of Gemini 4 Carbon had already shown Google's intention to spread its family across use cases, and Hydrogen completes the lower rung of that ladder.
Four thinking levels, labeled low, medium, high and xhigh, allow depth to be adjusted without changing model or rate. The high level is the default and xhigh is new, designed for when a problem requires more reasoning steps before answering. The company maintains that the progression is predictable, so the user pays in latency and reasoning tokens for what they need rather than for an average.
That design has a consequence for teams building agents. If the expensive and the cheap model share interfaces and flat prices, the natural move is to let an orchestrator decide which one to use at each step, instead of picking one for the whole workflow. Google already does this in Antigravity with Orchestrator, and that piece explains why Hydrogen exists: not to replace Carbon, but to save tokens on routine steps.
The API developer note adds the practical detail of immediate availability on the same day, November 12, with instructions for AI Studio and Android Studio and a reference to the four thinking levels as the only control the developer needs to touch.
Flat price across one million tokens
The flat rate is the launch's strongest commercial argument. In models with tiered pricing, a long conversation makes every new token more expensive because the context grows and is resent on every call. Hydrogen charges the same for token number one thousand and token number nine hundred ninety-nine thousand, and that linearity makes it possible to budget an agentic workflow without calculating how many times the model will reread its own history. The comparison reference is Gemini 3.8 Flash, released in September.
Almost every figure in the announcement is compared with that previous generation, which is the model Hydrogen replaces in the company's catalog. The figures in the following table are the ones Google chose to highlight.
| Metric |
Gemini 4 Hydrogen |
Reference |
| Input price |
$1.50 per million |
Same across the entire window |
| Output price |
$7.50 per million |
Same across the entire window |
| Context and output |
1 million input, up to 1 million output |
Generational leap |
| Terminal-Bench 4.1 |
57% |
18.2% in Gemini 3.8 Flash |
| Artificial Analysis index |
51 by default, 54 at xhigh |
50 in Gemini 4 Argon |
| Task speed |
2.5 times faster |
Versus Gemini 3.8 Flash |
One million output tokens is an unusual figure and should be read with care: it does not describe a normal conversation, but rather jobs that produce documents, reports or extensive code in a single go. There the cost soars at any rate, because output is billed at 7.50 dollars per million, and the flat-window saving only pays off if the work genuinely makes use of the long context.
The other side of flat pricing is that it reduces the temptation to split tasks up to save money. A team that previously broke an analysis into five short calls can now send it whole and pay the same per token, which simplifies the code and weakens discipline.
Where does it show in agent tasks?
The improvements the announcement highlights for agents have to do with behavior when something fails. Hydrogen adapts better when a tool call returns an error or when a search comes back empty, and the vendor puts the drop in fabricated claims at more than five times that of 3.8 Flash. In a flow with several chained tools, that detail is worth more than a few benchmark points, because a single invented step ruins the entire result.
Orchestrator, the piece Google debuts in Antigravity, switches models within the same session: Carbon takes on the steps that require judgment and Hydrogen executes the rest. On Terminal-Bench 4.0, the company puts the result at a success rate close to Carbon's with 24% less quota consumed and 19% less time. These are the vendor's figures about its own orchestrator, and they serve more as a description of the design than as proof of performance.
Availability is broad for a model that has just launched. Hydrogen arrives in AI Studio, the Gemini API, Android Studio, Gemini for business, the Gemini app, Google Search and Sheets for AI Pro and Ultra subscribers, and the availability detail by plan makes clear that free users get access to it when the automatic selector considers it the right model. In Search, it is the default model for AI mode.
For a team already paying for agents, the calculation is straightforward: if the small model handles most of the work, the cost per task falls. The case of Muse Ember and cheap models for agents shows the same logic at another vendor.
The benchmarks are supplied by whoever sells the model
The figure that will be cited most is the one from the Artificial Analysis index, where Hydrogen beats Argon by one point at its default level and reaches 54 at the xhigh level. One point on an aggregate index is not a significant difference in real-world use, and the announcement itself presents it alongside speed, which is where the advantage seems clearest.
The rest of the figures come from the same place. Terminal-Bench 4.1 at 57% versus the previous 18.2% and 58.8% on HLE-Diamond are the vendor's measurements. To these are added 71.6% on Agents' Last Exam and fifth place on Agent Arena, tests in which the vendor takes part. The partners' numbers, with runs between two and 2.8 times faster and between 12% and 23% fewer tokens, also carry no name or published methodology in the note.
That opacity is not exclusive to Google and does not invalidate the launch. It means the honest comparison will come in a few weeks, when independent evaluations appear and teams report what each task cost them with the previous model. Until then, the only figure that can be verified without intermediaries is the price, and that one is published.
The competition is not standing still either. The launch of the Mythos family and Anthropic's funding round show that rivals are also placing mid-tier models with aggressive pricing, and that pressure is what explains the flat rate.
Flat pricing as a sales argument
The underlying reading is that Google has stopped selling tokens and started selling finished tasks. With the flat window, the customer can think in terms of how many jobs they complete per month rather than how many calls they make, and the provider commits to long context not being a billing trap. It is a change of conversation that favors those with long agentic flows, precisely the customer DeepMind is interested in.
It remains to be seen whether the price holds when real workflows actually use the million input tokens on every call. The flat rate is a promise of predictability, and promises of that kind are measured on a customer's first full invoice under serious load. Meanwhile, Carbon and Argon remain the models for thinking, and Hydrogen is the one that executes.