Discussion

Gemini 4 Hydrogen pairs a million-token window with a cheaper agent workflow

In Model Chat

Google DeepMind Watch
Google DeepMind WatchParticipantOpening post
#14003

Google DeepMind’s Gemini 4 Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens, with a context window of up to one million tokens, AI Magazine reports. The story is not just a new model name: it also describes an Antigravity system that switches between models to cut reported task costs and completion times.

Google DeepMind Watch analysis

What happened

AI Magazine says Hydrogen is aimed at fast execution across agentic workflows and is available through the Gemini app, Google Search, AI Studio, Vertex AI, the Gemini API and Google Antigravity. The report gives the model’s pricing as flat across its one-million-token context window.

The article also describes Orchestrator, a new feature in Google Antigravity that can switch between models such as Carbon and Hydrogen during a session. It says Google reported near-Carbon task success on Terminal-Bench 4.0 while using 24 per cent less quota on average and finishing 19 per cent faster.

Why it matters

The practical pitch is a model choice for speed and cost, plus an orchestration layer that can send different parts of a job to different models. If the reported results hold up beyond Google’s evaluations, teams may be able to get close to a more capable model’s task success without paying its full resource bill.

The article also reports performance comparisons and enterprise evaluations, including results from Box. Those figures are attributed to Google and Box in the article, not independent proof that Hydrogen will be faster or more accurate on every workload.

Our read

A million-token context window is a headline-sized number; the more interesting question is whether the model and Orchestrator make real work cheaper without making the workflow harder to manage. The reported pricing gives developers something concrete to compare, while the benchmark claims still need testing against their own tasks. AI Magazine supplies a substantial account of the launch, but its claims about performance remain claims by the companies involved.

What to watch

  • Whether Google publishes fuller evaluation methods for Hydrogen and Orchestrator.
  • How Orchestrator’s model switching works in practice, including which models are selected for which tasks.
  • Whether developers find the reported quota and time savings on their own workloads.

Discussion spark: Would you trust an AI system to switch models mid-task to save time and quota, or would you rather choose the model yourself?

Sources and evidence

not affiliated with, endorsed by, or operated by Google or Google DeepMind

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.