Google ships Gemini 4 Hydrogen, a fast execution model that leaves the thinking to Carbon
On 12 November 2026, Google launched Gemini 4 Hydrogen, a fast and cheap model derived from Gemini 4 Carbon, built to write code and run agentic tasks while the frontier models plan. The new Orchestrator option in Antigravity turns that division of labour into a product feature: meter your reasoning tokens and spend them only where they earn their keep.
12 November 2026. Google launched Gemini 4 Hydrogen, a fast model adapted from Gemini 4 Carbon and aimed squarely at code execution and agentic workflows. Launch pricing sits at $1.50 per million input tokens and $7.50 per million output tokens — a long way below the frontier models. The same day, the Orchestrator option in Google Antigravity revealed the real story: a single project can now let Carbon plan and Hydrogen execute, switching models as the task evolves. Why it matters: the model race is no longer decided by the highest benchmark score alone, but by the routing between an expensive planner and a cheap executor — an architecture decision every team building an agent will now have to make for itself.
Two models, two gears: the Gemini 4 family specialises
Gemini 4 Hydrogen is not a straight successor to Gemini 3.8 Flash. It occupies a distinct role in the Gemini 4 lineup: Carbon and Argon carry the architectural judgement and extended reasoning, while Hydrogen handles fast execution — writing code, running tests, verifying changes, doing research, and driving tool-using agents. It accepts text, images, audio, and video, and its knowledge cutoff is March 2026.
The economics are explicit. Frontier models are expensive to serve, yet a large share of an agent’s work needs velocity rather than deep reasoning: applying a patch, rerunning a test suite, reading a documentation page. By isolating those steps in a cheaper model, Google charges for thinking at its true price and for execution at its own.
The model card gives measurable comparisons. Against Gemini 3.8 Flash, Hydrogen finishes tasks 2.5 times faster at the “high” thinking level, jumps from 18.2% to 57% on Terminal-Bench 4.1, and cuts fabricated claims more than fivefold. On the Artificial Analysis Intelligence Index, it scores 51 at high — above Argon’s 50 — and 54 at “xhigh”. The important qualification comes from the model card itself: on the cheating-adjusted Terminal-Bench 4.1, Claude Sonnet 5.5 still leads with 63.9% against Hydrogen’s 57.0%. In plain terms, Hydrogen improves sharply on the old Flash without beating every competitor on every comparison.
Orchestrator: planning-versus-execution becomes a product
The real news is not the model but the orchestration. Inside Google Antigravity, the Orchestrator option applies the division of labour within a single session. Carbon explores dependencies and produces a plan in a read-only phase; Hydrogen then edits code, updates packages, and runs tests. If execution hits friction, control returns to Carbon to inspect the state and repair the plan.
The difference from a static router is one of kind, not degree. A static router commits an entire task to one model based on the opening prompt; Orchestrator changes models as the work changes. The Antigravity team reports near-Carbon task success on Terminal-Bench 4.1 while using 24% less quota and finishing 19% faster on average. For parallel work, Carbon coordinates while Hydrogen powers worker and verifier agents in isolated workspaces.
Rollout follows the subscription ladder: Hydrogen reaches paid AI Pro and AI Ultra subscribers in Antigravity, while Orchestrator launches on AI Ultra, with enterprise access through Gemini for Business.
A one-million-token window, and four thinking levels
The token budget becomes a resource to manage deliberately. Hydrogen shares a single 1,048,576-token window across input, thinking, and output. Synchronous responses cap at 65,536 tokens per hop, with background execution or client-side continuation able to extend generation up to the remaining capacity of the window.
Hydrogen exposes four thinking levels: low, medium, high by default, and the new “xhigh”. Google recommends “high” for most complex tasks and reserves “xhigh” for harder multi-step problems where “high” falls short, noting that “xhigh” can overthink simpler prompts. Those settings make the shared budget consequential: reasoning consumes capacity that could have held input or final output, and higher settings also add latency. For a team paying per API call, that is one more cost variable to watch — and a reason not to leave “xhigh” as the default on a pipeline.
The model’s underside is documented honestly. Google reports reductions of up to 81% in agentic overreach and 79% in under-service versus Gemini 3.8 Flash, along with stronger resistance to indirect prompt injection — failure modes that matter the moment a model can pick tools and act across connected services. On the other side of the ledger, the model card records a 5.57-percentage-point rise in unjustified refusals versus Gemini 3.8 Flash, and lists hallucinations, occasional slowness, and higher token use at elevated thinking levels.
What it changes for anyone building agents
Hydrogen’s broader significance lies less in its scores than in the division of labour it makes affordable. Frontier reasoning can be reserved for the hard steps, while a fast execution model absorbs the volume — tests, edits, log reading, configuration generation — that in practice dominates an agent’s wall-clock time.
It is the same trade-off an SRE already makes between an expensive and a cheap worker in a queue: don’t pay premium rates for a task that doesn’t need them. What is new is that this split leaves the realm of in-house improvisation and becomes a product feature, with measured quota and latency guarantees. The central question of this launch remains: will the reported efficiency gains — 24% less quota, 19% less time — survive everyday use, once the documented refusals and reliability limits enter the picture? That will be answered in real pipelines, not launch documents.
The verdict
If your agent spends most of its time executing — generating code, running tests, editing files —, benchmark Hydrogen’s cost against your current frontier model: at $1.50 per million input tokens, the margin is substantial the moment execution dominates. Keep the frontier model for planning and hard debugging, and don’t leave “xhigh” switched on across a whole pipeline: reserve it for the multi-step problems where “high” genuinely fails.
If you are building a multi-step agent, the signal worth taking home is not the model but the Orchestrator architecture. An expensive planner producing a read-only plan, a cheap executor applying it, and a hand-back to the planner when execution stalls: that is a pattern you can reproduce with any pair of models, and it moves the question from “which model” to “which model for which step”.