It is two in the morning, and a developer is staring at a terminal. Her coding agent has been chewing through a bug for twenty minutes, and the bill is ticking up with every token it burns. She does not need a smarter model. She needs a faster one that does not cost a fortune.
Google thinks it has built exactly that. On November 12, the company launched Gemini 4 Hydrogen, a new member of its Gemini 4 family designed for one job: doing everyday work quickly and cheaply.
Hydrogen is not the flagship. That honour still belongs to Gemini 4 Carbon, the heavyweight Google unveiled two days earlier. Hydrogen is the workhorse — the model that runs the routine jobs, the long coding sessions and the multi-step agent tasks, while the big models wait for the hard problems.
What Is Gemini 4 Hydrogen?
Gemini 4 Hydrogen is an execution-focused AI model, adapted from Gemini 4 Carbon and tuned for speed rather than maximum capability. Google describes it as the model for everyday coding and agentic workflows — the kind of work where an AI writes code, runs tests, checks the results and tries again.
It accepts text, images, audio and video, and it can produce text in return. Its knowledge stops at March 2026, which matters less for coding than it would for current-affairs questions. What matters more is the size of its memory: a full one-million-token context window, with up to one million output tokens to match.
That window is shared across input, thinking and output, so long documents and large codebases fit in a single pass. Hydrogen can swallow dense PDFs, spreadsheets, charts, audio and recorded video, then turn them into polished, cited reports. For developers, that means whole repositories, not just single files.
Gemini 4 Hydrogen Speed: 2.5x Faster Than Gemini 3.8 Flash
Speed is the headline. According to Google, Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking setting. It also scores far higher on Terminal-Bench 4.1, a test of software agents working through a computer terminal: 57 percent, against Flash’s 18.2 percent.
It also makes things up less often. Google says Hydrogen cuts fabricated claims by more than five times compared with Flash, and it recovers cleanly when a tool call fails or a search comes back empty. Those are the small failures that usually turn a coding session into a mess.
The model offers four thinking levels — low, medium, high and a new xhigh tier. High is the default, and Google recommends it for most complex tasks. The xhigh setting is reserved for harder multi-step problems, with a warning that it can overthink simpler prompts. More thinking is not always better thinking.
Independent testing tells a more nuanced story. Artificial Analysis found that moving from high to xhigh lifted Hydrogen’s Intelligence Index score from 51 to 54, but more than doubled the estimated task time, from six minutes to thirteen, and raised the cost from $1.87 to $3.16 per task.
How Gemini 4 Hydrogen Scores on Benchmarks
On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high setting — just ahead of Gemini 4 Argon at 50 — and 54 at xhigh. That puts it in the middle of the frontier pack: capable, but not the model you reach for when only the best answer will do.
On agentic benchmarks the picture is stronger. Hydrogen ranks fifth on Agent Arena, leads its class on HLE-Diamond at 58.8 percent, and tops the Vals Tax Agent Bench outright at 52.3 percent. It ties for first on the Vals Legal Research Bench at 55.8 percent and leads its class on AutomationBench-AA.
For knowledge work over visual files, Google reports 53.4 percent on the Multimodal Knowledge Agent benchmark and 71.6 percent on Agents’ Last Exam v1.1. All of these figures are vendor-reported. The model card itself notes that Claude Sonnet 5.5 scores 63.9 percent on the cheating-adjusted Terminal-Bench 4.1 table, above Hydrogen’s 57 percent.
There are other caveats in the fine print. The model card records a 5.57 percentage-point rise in unjustified refusals compared with Gemini 3.8 Flash. And while Google’s safety assessment found no material capability increase over Argon, Hydrogen reached the alert thresholds for biological-information and cybersecurity risks, even though it stayed below the critical capability levels.
Gemini 4 Hydrogen Pricing: $1.50 Per Million Input Tokens
Hydrogen is priced flat across its entire one-million-token context window: $1.50 per million input tokens and $7.50 per million output tokens. That is a quarter of Carbon’s introductory price, and it undercuts most frontier models by a wide margin.
There is a catch worth knowing. Gemini 3.8 Flash currently costs $0.75 per million input tokens and $3.75 per million output tokens, but those rates rise to Hydrogen’s levels on January 1, 2027. The cheap tier is being folded into the new pricing, not preserved alongside it.
Google has also changed how repeated input is discounted. Hydrogen drops explicitly created context caches and uses automatic, implicit caching for prompts of at least 4,096 tokens. Cache hits cost $0.075 per million input tokens — a 95 percent discount. For agents that re-read the same codebase all day, that matters more than the headline price.
The economics only work if the model actually finishes the job. A cheap model that needs three attempts costs more than an expensive one that gets it right first time. That is why Hydrogen’s benchmark scores matter as much as its price tag.
Orchestrator: Pairing Gemini 4 Hydrogen With Carbon and Argon
Hydrogen is also built to work as part of a team. Google is introducing Orchestrator in Google Antigravity, a system that routes work across the Gemini 4 family: Hydrogen handles the routine execution, while Carbon and Argon take the harder reasoning.
According to Google, Orchestrator reached near-Carbon task success rates on Terminal-Bench 4.0 while using 24 percent less quota and finishing 19 percent faster on average. In Antigravity, Hydrogen powers the worker and verifier agents running in isolated workspaces, checking each other’s output before anything ships.
Orchestrator is rolling out first to paid Google AI Ultra subscribers, with enterprise access through Gemini for business. Hydrogen itself is rolling out to paid Google AI Pro and Ultra subscribers in Antigravity. The message is clear: Google wants the Gemini 4 family sold as a system, not as three separate models.
Where You Can Use Gemini 4 Hydrogen Today
Hydrogen launched on November 12 across Google AI Studio, the Gemini API, Gemini for business, Google Search and the Gemini app. Developers get free and paid tiers in AI Studio and the API, and the model is available to paid Pro and Ultra subscribers in Antigravity.
Free users of the Gemini app get Hydrogen too, but only through automatic model selection — they cannot pick it directly. That is a deliberate choice. Google is treating Hydrogen as the default engine for everyday work, the model that quietly handles the routine while users never see its name.
The launch lands in a crowded week for Google. Gemini 4 Carbon went generally available on November 10, Gemini Robotics 2.5 arrived the same day, and the Gemini Decisions API followed on November 11. Hydrogen is the fourth Gemini 4 announcement in three days.
What Gemini 4 Hydrogen Means — and What Comes Next
Hydrogen is not the model that wins headlines. It is the model that wins invoices. Google is betting that the next phase of the AI race is not about the smartest model, but about the cheapest model that is good enough — the one that can run a thousand small jobs a day without anyone worrying about the bill.
The test will be whether Hydrogen holds up outside Google’s own benchmarks. Independent scores so far are respectable rather than remarkable, and the xhigh trade-off shows the model’s limits. But at $1.50 per million input tokens, respectable may be enough.
