Gemini 4 Hydrogen brings near-Argon intelligence at $1.50 per million tokens, and now powers Search

Google DeepMind released Gemini 4 Hydrogen, a fast, efficient model for everyday coding and agent work, below the frontier Gemini 4 Carbon (our analysis) and Argon models. It is rolling out globally across AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search and Google Antigravity for paid Pro and Ultra subscribers; free Gemini app users get it through automatic model selection.

Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default setting, edging past Argon’s 50, at a flat $1.50 per million input tokens and $7.50 per million output tokens across its full 1 million-token window. It offers four thinking levels, low, medium, high (the default) and a new xhigh, where Google says it reaches 54. It accepts up to 1 million input tokens and can produce up to 1 million output tokens. Google also cites 58.8% on HLE-Diamond, first place on Vals Tax Agent Bench at 52.3% and 96.7% on its internal agent honesty evaluation.

The interesting shift isn’t that a smaller model got smarter. It is that Google is openly designing for a frontier model that plans and a cheaper one that executes. Antigravity’s new Orchestrator switches between Carbon and Hydrogen within a session; Google says that on Terminal-Bench 4.0 it reached near-Carbon success while using 24% less quota and finishing 19% faster. If that holds in real teams, agentic coding becomes affordable well beyond big companies. It mirrors OpenAI’s cheaper GPT-6.1 Terra push (our analysis) and Google’s own decision-only model for agents (Decisions API).

Against Gemini 3.8 Flash, Google claims 2.5x speed at default, Terminal-Bench 4.1 up from 18.2% to 57%, and fabricated claims cut more than five times. Partners reportedly saw 2x to 2.8x speedups with 12% to 23% fewer tokens. These are Google’s figures and early partner reports, strong claims awaiting independent confirmation.

Google says Hydrogen cuts agentic overreach by up to 81%, reduces misaligned actions and task shortcutting 12-fold, is up to 2x less likely to bypass permission boundaries on blocked tasks, and is more than 2.3x more resilient on Gray Swan’s indirect prompt injection benchmark. Those are the right things to measure for a model that reads inboxes and calls tools. But “up to 2x less likely to bypass permissions” still means it sometimes does, and a small failure rate multiplied by a global user base is many incidents.

Hydrogen now powers Search for all users, so it is effectively Google’s default intelligence for everyday queries (the Search upgrade). Google says it broadens query fan-out to surface a wider range of links; publishers will want to know whether that means more traffic or more citations inside an answer. In the same week families protested Google’s AI in schools (coverage), it is worth remembering a model in Search reaches every user, including the youngest. Overall, this is a strong, practical release that pushes capable agents down the price curve; the test is whether its safety gains hold outside Google’s evaluations. More in our Google Gemini coverage.