A developer working on a laptop in a bright office, code on the screen, photographed in natural light

On Thursday, November 12, Google released Gemini 4 Hydrogen, the third model in its Gemini 4 family — and the first one aimed at everyone. Argon arrived September 30 and Carbon on October 28, both positioned as heavy reasoning engines. Hydrogen is the workhorse: built for coding, fast tool use, and the many smaller tasks that dominate a multi-agent workflow. It is also, from today, the model answering the hardest questions in Google Search.

Hydrogen rolls out across Google AI Studio, the Gemini API, Google’s business products, the Gemini app, and Search — where it becomes the default model in AI Mode and the model behind the hardest questions in AI Overviews. A frontier model launch that lands in Search on day one is not a developer-only event. It is a change to what billions of people see when they ask Google a question.

What Google says it can do

The numbers Google published are, as always, vendor-reported — but the shape of them is the story:

  • Price: $1.50 per million input tokens and $7.50 per million output tokens, across a one-million-token context window, with four thinking levels: low, medium, high, and xhigh.
  • Coding: In Google’s Terminal-Bench 4.1 comparison, Hydrogen scored 57%, against 18.2% for Gemini 3.8 Flash — the previous fast model in the lineup.
  • General capability: Google reports a default high-thinking score of 51 on the Artificial Analysis Intelligence Index, against 50 for the far pricier Gemini 4 Argon. If that holds up independently, the cheap model is roughly matching the flagship on general intelligence — which is exactly the point of a model like this.
  • Routing: Google also announced Orchestrator inside Antigravity, its agentic coding environment, which shifts work dynamically between frontier and lower-cost models. Google says Orchestrator used 24% less quota and finished 19% faster on Terminal-Bench 4.0 than its comparison setup, while retaining near-Carbon task success.

The division of labor is the real product here. One stronger planner coordinating cheaper workers has been the standard way serious teams run agents for a year; Hydrogen puts a concrete commercial price on it. The question for teams is no longer whether to split work across models, but whether Google’s routing does it better than their own.

The sibling launch: decisions as a product

Hydrogen did not arrive alone. A day earlier, on November 11, Google launched the Gemini Decisions API, built around a model called Gemini Decision 1. It does not generate free-form text. It scores bounded choices — yes/no, multiple choice with 2 to 255 candidates, ordered rubrics — over text, image, audio, video, or PDF input, returning option probabilities and confidence scores instead of a paragraph that has to be parsed back into a decision.

The price is $0.10 per million input tokens with no output-token charge, and Google quotes 150 to 250 milliseconds end-to-end latency. The use case is the unglamorous middle of agent work: which tool to invoke, whether a document meets a rule, whether a human reviewer should intervene. A specialized primitive like this could cut latency and cost where agents currently burn a full generation on a one-word answer — but it is also a new dependency with its own failure modes, and a confidence score of 0.8 is still wrong one time in five.

The day in context

November 12 was one of the busiest single days in AI this year, and Hydrogen landed in the middle of it:

  • Anthropic opened a private beta of self-hosted Claude Managed Agents, letting enterprises run the whole agent service — harness, sandboxes, credentials, session history — inside their own cloud or on-premises. The model itself stays with the provider; the control plane moves into the customer’s house.
  • OpenAI launched Daybreak Red Team and Defense Factory, security agents that penetration-test web apps and draft fixes a human must still approve, built on its restricted GPT-6 Cyber model.
  • Four nonprofits — the Irreplaceable Education Fund, Public Citizen, the Heat Initiative, and the Alliance for Secure AI — sued OpenAI, Anthropic, Google, Meta, and SpaceX’s AI division in a San Francisco court, seeking an injunction to freeze development of any AI system more capable than the labs’ current ones. No hearing date has been set, and none of the companies has responded publicly.
  • SpaceX outlined plans for more than $74 billion of AI supercomputing campuses in Texas and Arizona, and AMD announced the Instinct MI450X accelerator with 288 GB of HBM4 memory.

Read together, the day had a shape: the frontier is being split into cheaper, faster pieces, the control plane is moving toward the customer, and the courts are being asked to decide whether any of it should continue at all.

What to watch

Three things will tell us whether Hydrogen matters as much as its launch day suggests. First, independent benchmarks at matched effort settings — Google’s Terminal-Bench and Intelligence Index figures are its own, and the gap between Hydrogen and Argon is small enough that independent testing could erase or reverse it. Second, real cost per completed task: a million-token window does not make a million-token prompt cheap, and routing mistakes in Orchestrator could erase the apparent token savings. Third, whether Search users notice — Hydrogen’s hardest test is not a benchmark but a billion people asking it questions it was not tuned for, in a product where a wrong answer is visible immediately.

Google has now shipped three Gemini 4 models in six weeks. The pattern is no longer a single flagship with a cheaper sibling. It is a family, priced by the job — and the cheapest member just became the one most people will actually meet.