Google has launched Gemini 4 Hydrogen, a fast new AI model for coding and everyday work, rolling it out on Thursday to developers, businesses and free users of its app and search.

The launch completes a rapid run of releases from Google. Hydrogen is the third model in the Gemini 4 family to ship in just over six weeks, after Argon and Carbon, and it takes over the job the company’s Flash models used to do: quick, cheap, everyday work. Unlike its two bigger siblings, it is not locked behind a paywall — at least not entirely.

Key Points

  • Gemini 4 Hydrogen launched on 12 November, the third Gemini 4 model in just over six weeks, after Argon and Carbon.
  • Google pitches it as the fast execution model for coding and AI agents, replacing the old Flash branding.
  • It costs $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1-million-token window.
  • At its default setting it scores 51 on the Artificial Analysis Intelligence Index, one point ahead of the flagship Argon.
  • Free users get it only when the Gemini app’s automatic selection routes a prompt to it; paying users can pick it directly.

What happened

Google announced the model on Thursday in a blog post from Koray Kavukcuoglu, the head of Google DeepMind and the company’s chief AI architect. He described Hydrogen as built for rapid execution across agentic workflows — the kind of work where an AI writes code, calls tools, checks its own output and tries again, sometimes hundreds of times in a single job.

The model is rolling out globally from today across Google’s developer, business and consumer products: the Gemini API, Google AI Studio on free and paid tiers, the Gemini app, Google Search, Gemini for business, and Antigravity, the company’s coding platform, for Pro and Ultra subscribers. Google says it is also coming to Gemini in Sheets for Pro and Ultra subscribers.

Hydrogen is the fast end of the Gemini 4 family. Carbon and Argon, the two models Google released first, are built for deep, long-horizon reasoning on the hardest problems. Hydrogen is the one meant to do the ordinary work quickly: writing code, running tests, fixing bugs, reading documents and answering questions. It takes text, images, audio and video as input and produces text, and it can handle up to 1 million tokens of input at a time, with up to 1 million tokens of output.

It also brings a new control. Hydrogen has four thinking levels — low, medium, high and a new xhigh tier — so developers can choose how much reasoning a task gets, and pay only for what the job needs. The default is high, a step up from the medium default on the old Flash model.

Google is also building the choice between models into Antigravity. A new feature called Orchestrator switches between Carbon and Hydrogen inside a single session, putting the bigger model on planning and the faster one on execution. Google says the combination came close to Carbon’s success rate on its Terminal-Bench 4.0 test while finishing 19% faster and using 24% less quota.

Why it matters

The numbers Google puts forward are strong, and they are all the company’s own. On the Artificial Analysis Intelligence Index, an independent test that combines nine evaluations, Hydrogen scores 51 at its default setting, one point ahead of Argon and well ahead of the 31 scored by Gemini 3.8 Flash. At its highest thinking level it reaches 54. Google says it finishes tasks 2.5 times faster than Flash, and on Terminal-Bench 4.1, a test of work done through a computer terminal, it completes 57% of tasks against 18.2% for Flash. It still trails Anthropic’s Claude Sonnet 5.5, which scores 63.9% on the same test, according to VentureBeat.

The price is where the launch gets complicated. At $1.50 per million input tokens and $7.50 per million output tokens, Hydrogen costs twice as much as Flash’s introductory rate, which runs until 31 December. From 1 January, Flash’s own standard price rises to exactly the same $1.50 and $7.50. VentureBeat notes that Hydrogen costs about 15 times as much per token as the cheapest rates for Anthropic’s Claude Haiku 5.5, OpenAI’s GPT-6 Luna and Meta’s Muse Ember — though those rivals score lower on the same broad index. Cheaper batch and flex tiers, at half the standard rates, are available for work that can wait.

Early testers have seen real gains. Box, the cloud storage company, found its AI agent’s accuracy rose from 74% to 79% when it switched from Flash to Hydrogen, while taking less than half as long and using about 12% fewer tokens, its chief executive Aaron Levie said on X. The biggest improvements came on work that involved reading tables and doing calculations.

The free tier comes with a catch that Google’s announcement does not spell out. Free users of the Gemini app cannot choose Hydrogen from the model menu — only paying subscribers can. Free users reach it only when the app’s automatic selection decides a prompt needs it, and most automatic prompts go to the lighter Flash-Lite model instead. In Search, though, the picture is different: Google says Hydrogen is now the default model in AI Mode and is used for the hardest questions in AI Overviews, so free users will meet it there whether they pick it or not.

Google also published safety figures for the model, again from its own testing. It says Hydrogen cuts overreach in agent tasks by up to 81% and under-service by up to 79% compared with Flash, and that it is more than twice as resilient to indirect prompt-injection attacks. It also says the model invents facts more than five times less often than Flash. None of those figures has been checked by an outside tester yet.

What happens next

The real test is the one Google’s own launch hints at: whether teams can run their everyday agent work on a fast model and keep the expensive ones for the jobs that truly need them. Orchestrator, the tool that splits work between Carbon and Hydrogen automatically, is the clearest sign of where Google thinks the answer lies — not one model for everything, but the right model for each step.

For developers still on Flash, the clock is ticking in a quieter way. There is no shutdown date for the old model, but its price doubles on 1 January to match Hydrogen’s, which removes the last reason to stay. The migration is not just a name change: Hydrogen rejects the old thinking-budget setting and starts automatic caching at a lower threshold, so teams will need to adjust their code.

The rivalry is not standing still either. Carbon, the strongest model in the family, opened to paying developers two days before Hydrogen, and Anthropic’s Claude Fable 5.2 arrived the day before that. Google’s Decisions API, launched on Tuesday, handles the small yes-or-no choices inside an agent — the kind of job Hydrogen is now built to do at scale. The race is no longer about who has the smartest model. It is about who can do the everyday work fastest, cheapest and most reliably.

Sources: Google blog, 12 November 2026; VentureBeat, 12 November 2026; 9to5Google, 12 November 2026; SiliconANGLE, 12 November 2026; AndroidPure, 13 November 2026. Published 13 November 2026.