Google has a new workhorse for the age of AI agents. On November 12, 2026, the company introduced Gemini 4 Hydrogen, a fast, low-cost model built for the everyday work that coding agents actually do: writing code, running tests, verifying changes, doing research and calling tools in a loop. It is the execution engine of the Gemini 4 family — the model you reach for when the job is doing, not deliberating.
Hydrogen is adapted from Google’s Gemini 4 Carbon, the long-horizon reasoning model Google launched earlier this month, but it plays a deliberately different role. Carbon and Argon handle the hardest reasoning; Hydrogen handles the volume. As Google’s announcement puts it, Hydrogen brings “cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows.” That positioning matters because agent workloads are not one big question — they are hundreds of small ones, and the per-token price of each one decides whether an agent loop is affordable at all.
What Gemini 4 Hydrogen Is Built For
Hydrogen is designed for rapid execution across agentic workflows. In practice that means issue resolution, debugging, rapid iteration, codebase question-and-answer, and the focused, surgical code edits that keep a project moving without rewriting unrelated files. Google says it delivers “cleaner, production-ready code” rather than sprawling rewrites — a direct answer to one of the most common complaints about AI coding tools.
It is also built to be a team player. Hydrogen can stand on its own as the model for everyday work, or it can pair with Carbon and Argon as the high-efficiency execution engine in multi-model workflows. That division of labor is the story of this launch as much as the model itself: Google is no longer selling one model for everything, but a family where each member has a job. It is the same direction the industry is moving — Google’s Gemini Decisions API, launched just a day earlier, handles bounded yes/no and scoring decisions for agents, while always-on assistants like Databricks Genie show how agents are becoming permanent coworkers rather than one-off tools.
Beyond code, Hydrogen is a multimodal workhorse. It accepts text, images, audio and video, and pairs a native 1 million-token input window with up to 1 million output tokens to synthesize dense PDFs, spreadsheets, charts, audio and recorded video into polished, cited reports and structured deliverables. One caveat to note: its knowledge cutoff is March 2026, so it may be less current in fast-moving domains.
Pricing That Makes Agent Loops Affordable
The headline number is the price. Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1 million-token context window — no long-context surcharge. That flat pricing is deliberate: agent loops fire hundreds of small requests per task, and unpredictable pricing kills them.
There are two more pricing details worth knowing. First, Hydrogen drops explicitly created context caches in favor of automatic, implicit caching: repeated prompts of at least 4,096 tokens are cached automatically, and cache hits cost $0.075 per million input tokens — a 95% discount. Second, Hydrogen takes over from Gemini 3.8 Flash as the default model for high-volume work. Flash stays at $0.75/$3.75 per million tokens through December 31, 2026, then rises to Hydrogen’s rates on January 1, 2027. A rate-limited free tier is available in AI Studio and the Gemini API for developers who want to try it first.
The economics show up in independent testing, too. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level — just ahead of Argon at 50 — and 54 at the new xhigh setting, at an estimated $1.87 per test task at high and $3.16 at xhigh. Speed is part of the value: at high, Hydrogen completed tasks faster than most models in its class, at about six minutes per test task.
Four Thinking Levels — Including a New xhigh Tier
Hydrogen offers four thinking levels — low, medium, high (the default) and a new xhigh tier — that scale quality, latency and cost in a smooth, predictable progression. The idea is simple: you only wait and pay for the reasoning depth a task actually requires. A routine code edit runs at low or medium; a tricky debugging session gets bumped to high or xhigh.
That xhigh tier is the genuinely new control. It lets a cheap default model spend more compute on the rare request that actually needs it, instead of forcing you to switch to a bigger, pricier model. The trade-off is real — moving from high to xhigh roughly doubles the estimated task time, from about six minutes to thirteen in Artificial Analysis testing — but for the hard 5% of tasks, that is a better deal than paying frontier prices on the other 95%.
One benchmark note keeps the picture honest. On Terminal-Bench 4.1, with cheating-adjusted scoring, Hydrogen reaches 57.0%, below Claude Sonnet 5.5 at 63.9%. Hydrogen is not trying to top the coding charts; it is trying to be the model you can afford to run a thousand times a day.
Orchestrator: Carbon Plans, Hydrogen Executes
The launch also brings Orchestrator to Google Antigravity, the company’s agent coding platform, and it is the clearest expression of the multi-model idea. Orchestrator dynamically coordinates models within a single session: for workflows that need parallel agents, Carbon can act as the coordinator while Hydrogen powers worker and verifier agents that build and test changes simultaneously across isolated workspaces.
Hydrogen is also now the default agent in Gemini Managed Agents inside Antigravity. Orchestrator is rolling out initially to Google AI Ultra subscribers, with enterprise access through Gemini for business, while Hydrogen itself is available to paid Google AI Pro and Ultra subscribers in Antigravity. The pattern is familiar from the wider agent market — NVIDIA’s OpenShell and the governance tooling from Commvault and Zscaler point the same way: agents are becoming systems of coordinated models, and the winners will be the ones who make that coordination automatic rather than manual.
Where You Can Use Hydrogen Today
Hydrogen is rolling out globally starting today across Google’s developer, enterprise and consumer surfaces: Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and Google Antigravity for paid Pro and Ultra subscribers. Developers can build with it in AI Studio and Android Studio, and free Gemini app users will receive it through automatic model selection when Google considers it appropriate for a prompt.
For developers, the model ID is gemini-4-hydrogen. It supports function calling, structured outputs, code execution, file search, Google Search grounding and the Interactions API, so it slots into existing Gemini-based agent stacks without a rewrite. Google also published a model card with safety assessments: targeted tests placed Hydrogen below the biological-information and cybersecurity critical capability levels, while reaching the alert thresholds in both domains — worth knowing for teams deploying agents in sensitive environments.
What to Watch Next
Hydrogen’s launch raises three questions that will decide how far it goes. First, does the multi-model pitch actually save money in production, or does the overhead of coordinating Carbon and Hydrogen eat the savings? Second, will the xhigh tier become the default for serious coding work, quietly pushing real-world costs above the headline price? And third, how will rivals respond — a fast, cheap execution model with a million-token window and a 95% cache discount is a direct challenge to every coding-agent provider charging frontier prices for routine work.
The broader signal is that the model market is splitting in two. There are the frontier reasoners for the hardest problems, and there are the execution engines for everything else. Hydrogen is Google’s bet that the second category is where most of the volume — and most of the value — will live.
Conclusion
Gemini 4 Hydrogen is not the most powerful model Google has ever shipped, and it is not trying to be. It is the most practical: fast enough for agent loops, cheap enough to run all day, smart enough to score 51 on the Artificial Analysis Intelligence Index, and flexible enough to scale its thinking from low to xhigh when a task demands it. Combined with Orchestrator in Antigravity, it turns the Gemini 4 family from a set of models into a system — Carbon plans, Hydrogen executes. For developers building coding agents, that system is available today, and the price makes it hard to ignore.
