Introducing Gemini 4 Hydrogen

Google DeepMind released Gemini 4 Hydrogen, a fast, efficient model meant for everyday coding and agent work, sitting below the frontier Gemini 4 Carbon 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.

Two things stand out. 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. And because it now powers Search for all users, it is effectively Google's default intelligence for everyday search, rolling out globally.

The interesting shift here is not that a smaller model got smarter. That happens every few months. It is that Google is openly designing for multi-model work, where an expensive frontier model plans and a cheaper one executes. Its new Orchestrator in Antigravity switches between Carbon and Hydrogen within a session, and Google says that on Terminal-Bench 4.0 it reached near-Carbon success while using 24% less quota and finishing 19% faster. If those numbers hold in real teams, that is the kind of efficiency that makes agentic coding affordable outside big companies. We think that is genuinely useful.

The upgrade claims against Gemini 3.8 Flash are large: 2.5 times faster at the default setting, Terminal-Bench 4.1 up from 18.2% to 57%, and fabricated claims cut by more than five times. Partners reportedly saw 2x to 2.8x speedups with 12% to 23% fewer tokens. These are all Google's own figures and early partner reports, so we treat them as a strong claim awaiting independent confirmation, not settled fact.

The safety section is where we spent the most time, because reach changes the stakes. Google says Hydrogen cuts agentic overreach by up to 81%, reduces misaligned actions and task shortcutting by 12 times, is up to 2x less likely to bypass permission boundaries on blocked tasks, and is more than 2.3 times more resilient on the Gray Swan indirect prompt injection benchmark. Those are the right things to measure for a model that will be reading inboxes and calling tools. But up to 2x less likely to bypass permissions still means it sometimes does, and a small failure rate times a global user base is a lot of incidents.

The Search piece deserves its own scrutiny. Google says Hydrogen broadens query fan-out to surface a wider range of links. Publishers will want to see whether that means more traffic or just more citations inside an AI answer. And in the same week families protested Google's AI in schools, it is worth remembering that a model deployed into Search reaches every user, including the youngest ones.

Our overall read: a strong, practical release that pushes capable agents down the price curve. The test is whether its safety gains hold up outside Google's own evaluations.

Hydrogen offers four thinking levels: low, medium, high (the default) and a new xhigh tier, where Google says it reaches 54 on the Artificial Analysis index. It accepts up to 1 million input tokens and can produce up to 1 million output tokens. Free Gemini app users get it through automatic model selection when appropriate. Google also cites results including 58.8% on HLE-Diamond, first place on Vals Tax Agent Bench at 52.3%, and a 96.7% rate on its internal agent honesty evaluation.