On 12 November 2026 Google released Gemini 4 Hydrogen, the fast variant of the Gemini 4 family built for everyday coding and agentic work. On the Artificial Analysis Intelligence Index Hydrogen scored 51 points, one more than the stronger and pricier Gemini 4 Argon, at a flat $1.50 per million input tokens across the entire one-million-token window. The model reached the Gemini API, the Gemini app and Google Search globally on the same day.
Key takeaways
- 51 points on the Artificial Analysis Intelligence Index at the default high thinking level, 54 points at the new xhigh level
- Flat pricing across the whole one-million-token window: $1.50 per million input tokens, $7.50 per million output tokens
- Terminal-Bench 4.1: 57 percent versus 18.2 percent for Gemini 3.8 Flash, with tasks finishing 2.5 times faster
- A new Orchestrator option in Google Antigravity switches between Carbon and Hydrogen inside one session — on average 24 percent less quota consumed and 19 percent less time
- Available globally from 12 November 2026 in the Gemini API, AI Studio, the Gemini app and Google Search
Let us picture the situation
A developer opens a web application repository in Google Antigravity and types a single instruction. The instruction reads, for example:
Update the outdated dependencies, fix whatever breaks afterwards, and run the tests.
A task like this has two entirely different phases. First someone has to work out what depends on what in the project and in which order to touch it — analytical work where the strongest available model pays off. Then a few dozen repetitive edits have to be made and the tests run, and here pace and discipline matter most.
But a problem appears: who is to decide which model handles which phase, when the boundary between them only emerges mid-task?
This is precisely the problem Orchestrator is meant to solve — a new entry in the model list in Google Antigravity that switches models in the middle of a single session. (The instruction above is an editorial example, not a quote from the source.)
What exactly is Hydrogen within the Gemini 4 family?
At Google, Gemini 4 is not one model but a family with separated roles. Carbon and Argon are the frontier?frontier: the strongest and most expensive model class a vendor offers, tuned for the hardest reasoning variants, tuned for deep reasoning on the hardest problems. Hydrogen sits beside them as the fast execution model — meant to write code, run tests and verify changes rather than design system architecture.
The comparison produces a puzzling result. At its default thinking level Hydrogen scores 51 points on the Artificial Analysis Intelligence Index against Argon's 50, and reaches 54 at the highest xhigh level. A model positioned as the cheaper and faster option therefore comes out ahead of one of the two frontier variants on this single aggregate index. Google does not explain the discrepancy beyond stating that Carbon and Argon retain the advantage on tasks requiring long-horizon reasoning.
| Variant | Role in the family | Artificial Analysis Intelligence Index |
|---|---|---|
| Carbon | frontier — deep reasoning | not disclosed |
| Argon | frontier — deep reasoning | 50 |
| Hydrogen | fast execution: code, tests, verification | 51 (high) / 54 (xhigh) |
Pricing is flat across the entire context window: $1.50 per million input tokens and $7.50 per million output tokens, with no step-up once part of the window is consumed. Hydrogen takes one million tokens of input and can generate up to one million tokens of output.
How does Orchestrator work?
Orchestrator is not a separate model but an entry in the model picker in Google Antigravity that combines Carbon and Hydrogen in a single run. Carbon starts in a read-only phase — mapping dependencies, resolving ambiguities and building a plan with explicit triggers for when the plan should be rebuilt. Execution then passes to Hydrogen: it edits files, updates packages and runs tests. If Hydrogen hits friction, control hands back to Carbon, which inspects the project state, repairs the plan and passes execution onward.
The diagram shows how a single Google Antigravity session splits into planning on Carbon's side and execution on Hydrogen's, and at which point control hands back to the planner. This is the flow Google described as an example, not a full specification of the switching rules.
The result reported by the Antigravity team: on Terminal-Bench 4.1 Orchestrator reaches task success rates close to Carbon alone while consuming on average 24 percent less quota and finishing 19 percent faster.
Orchestrator arrives in two variants — Orchestrator (Carbon) and Orchestrator (Argon) for Google AI Ultra subscribers, and Orchestrator (Argon) for paid Google AI Pro.
Why is a plain model router not enough?
The standard industry approach is a static router: the system grades task difficulty from the first prompt and routes the whole task to the cheapest model that can finish it. For short queries this works.
The Antigravity team points to where the approach falls apart. The longer an agent runs, the less the first prompt says about what will be needed at step thirty. Difficulty is not a property of the task but of its successive stages, and it surfaces only along the way. A router that decided once has no way to correct itself: either it pays the frontier rate for simple edits throughout, or it gets stuck on the cheap model at the exact moment the plan needs rethinking.
Orchestrator is bidirectional and switches models both ways within a single trajectory?trajectory: the full chain of steps an agent walks from the prompt to the result. That addresses a second problem a router never touches: context loss. Handing a task from model to model halfway through requires the project state and the plan to be recorded in a form the second model can read without starting over — which is why Carbon builds a plan with explicit rebuild triggers rather than just a list of steps.
The alternative is to assemble such a setup by hand through the Antigravity SDK, where the developer assigns the roles directly: Hydrogen agents for issue triage and fast iteration, a Carbon agent for larger refactors and security checks.
What do these names mean?
Four terms recur in Google's materials, and without them the numbers say little.
- Thinking levels (low, medium, high, xhigh) — a setting for how much reasoning the model performs before answering. The equivalent of a quality–time–cost slider. High is the default, xhigh is the new top level.
- Artificial Analysis Intelligence Index — an aggregate index from the independent Artificial Analysis service that folds results from more than a dozen tests into one number. It allows a single-figure comparison but blurs the differences between task types.
- Terminal-Bench — a test of whether a model can carry a terminal task through to completion: run commands, react to an error, verify the result. It measures execution, not knowledge.
- Indirect prompt injection — an attack in which a malicious instruction is hidden inside data the model fetches itself, for instance in page content or a file. The equivalent of code injection, except the vector is text.
Together these four terms describe what Google is selling in Hydrogen: not peak intelligence, but predictable execution cost with retained resistance to tasks that try to push an agent out of bounds.
Is it working already?
Yes, since 12 November 2026, globally and in production. Hydrogen is available across all tiers of Google AI Studio and in the Gemini API, in Gemini for business, in the Gemini app model picker, and in Google Search for all users.
Access is tiered, however. In Google Antigravity Hydrogen reaches paid Google AI Pro and Ultra subscribers only. Gemini in Google Sheets starts with AI Pro and Ultra subscribers and selected business plans. Orchestrator is the most restricted: the Carbon variant is limited to the Ultra plan.
Why does it matter?
Hydrogen moves the reference point in the price list: a model that beats one of the frontier variants on an aggregate index costs what the execution tier used to cost. Orchestrator is the more interesting part, though. In it Google concedes outright that one model for an entire agentic task is wasteful, and that model selection is a decision taken repeatedly during the work rather than once at the entry point.
What next?
- Orchestrator launches with a limited line-up — Google says more models and subscription plans will be added, without giving a date
- Gemini 3.8 Flash stays in the catalogue as the predecessor Google measures Hydrogen against (57 percent to 18.2 percent on Terminal-Bench 4.1), which points to migration as the intended path for teams on Flash
- The full list of limitations and results sits in the Gemini 4 Hydrogen model card published alongside the launch
Sources
- blog.google — Introducing Gemini 4 Hydrogen
- Google Antigravity Blog — Gemini 4 Hydrogen and Orchestrator
- blog.google — Google Search gets a major upgrade with Gemini 4 Hydrogen
- Ars Technica — Google rolls out Gemini 4 Hydrogen, finally replacing 3.8 Flash





