RecodeAI Explore AI Intelligence

Frontier Models & AI Economics

Gemini 4 Hydrogen and the rise of the router

Google stopped shipping a best model and started shipping a lineup — then put the thing that chooses between them inside its own platform.

2 min read Editorially reviewed 22 NOV 2026

In brief

Google released Gemini 4 Hydrogen around 12 November, built for fast, low-cost coding and agentic work rather than heavy reasoning. It sits in a four-model family alongside Carbon, Argon and Neon, with Antigravity Orchestrator routing between them. Google reports intelligent routing cutting task completion time by 19%.

What happened

Google released Gemini 4 Hydrogen in mid-November — most sources give 12 November — positioning it for fast, low-cost coding and agentic workflows rather than heavyweight reasoning. In the consumer app it is reached through automatic routing on the free tier, while paid subscribers can select it directly.

Hydrogen completes a differentiated family alongside Carbon, Argon and Neon, the last reported around 18 November as a mid-tier option. The routing layer, Antigravity Orchestrator, is described as a multi-model orchestrator; Google reports intelligent routing improving overall task completion time by about 19%. Google has also said it is bringing Hydrogen to Search globally. Reports describing it as the default model for AI Mode are not confirmed by Google's own pages, and third-party pricing figures vary enough that we have omitted them.

Why it matters

For three years the industry has argued about which model is best. Google's answer is that the question is malformed: a serious workload uses several, and the valuable position is owning the thing that decides which one runs.

That is a strategically different place to stand. A router sits between the customer and every model, including models that are not yours. It accumulates the data about which tasks need what — and that data, not any single checkpoint, is what compounds.

It also reframes what a model release means. For most of the last three years a launch was an attempt to claim the top of a leaderboard. A release positioned explicitly below the frontier, optimised for cost and latency, is an admission that the leaderboard was never the product — the workload was.

Business and industry implications

Routing is where enterprise AI economics are decided. Most organisations are paying frontier prices for a workload that is mostly routine, because standardising on one capable model is operationally simpler than managing several. A good router removes that trade-off — and moves the margin to whoever operates it.

The strategic caution is the obvious one. A platform-owned router is convenient and makes model choice someone else's decision. Enterprises that care about portability should understand whether they can inspect, override or replace the routing policy, or whether it is simply applied to them.

Leadership perspective

Intelligent routing is likely to become necessary to making enterprise AI economically sustainable — the spread between tiers is too large to ignore at volume. The question is whether the router is a capability you hold or a dependency you accept. Those look identical while it works well.

Three questions for business leaders

  1. What share of our AI spend runs on frontier models doing non-frontier work?
  2. Can we inspect and override routing decisions, or are they applied to us?
  3. Is routing a capability we should own rather than consume?