Google launches Gemini 4 Hydrogen, a fast model for everyday agents

The rollout is global. Google says it covers all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search, plus paid Google AI Pro and Ultra subscribers in Antigravity[1]. Developers can call the API from AI Studio and Android Studio[1][2]. Enterprises can deploy it in Gemini for business, through the API or the model drop-down, in Antigravity, and in Gemini in Google Sheets on an eligible business subscription[1]. The Gemini app lists it in the model drop-down. Search has it for all users. Sheets for individuals starts with AI Pro and Ultra. A footnote says free Gemini app users get Hydrogen through auto model selection, when it is the appropriate model for the prompt[1][3].

Hydrogen can run on its own, across four thinking levels, or beside Carbon and Argon[2]. The levels are low, medium, high (the default) and a new xhigh tier. Google says a task only waits, and only pays, for the reasoning it needs[1].

DeepMind's model card says Hydrogen is adapted from Carbon. It takes text, images, audio and video in a combined window of up to 1 million tokens. A synchronous reply stops at 64K output tokens, and a background long decode can use up to 1 million. The knowledge cutoff is March 2026. The card notes hallucinations, occasional slowness or timeouts, and higher token use at the upper thinking levels[4].

Google says Hydrogen is a major upgrade from Gemini 3.8 Flash. At the default high setting, it says Hydrogen finishes tasks 2.5 times faster, more than triples Terminal-Bench 4.1 (57% against 18.2% for 3.8 Flash), and cuts fabricated claims by more than five times. It also says the model adapts when a tool call fails or a search comes back empty[1].

Orchestrator, new in Antigravity, switches between models such as Carbon and Hydrogen inside one session. On Terminal-Bench 4.0, Google says it reached near-Carbon task success while using 24% less quota on average and finishing 19% faster. Where agents run in parallel, Carbon can coordinate and Hydrogen can power worker and verifier agents in isolated workspaces[1].

These figures are Google's own claims. On the Artificial Analysis Intelligence Index, Google says Hydrogen surpasses Argon at the default high thinking level, 51 to 50, and reaches 54 at xhigh[1]. The launch post also says Hydrogen leads its class on HLE-Diamond at 58.8%[1].

The November 2026 model card sets Hydrogen beside Claude Sonnet 5.5 and GPT-6.1 Terra. SWE-sweep is 13.5%, against 3.1% and 5.8%. Terminal-Bench 4.1, marked cheating-adjusted, is 57.0% for Hydrogen and 63.9% for Sonnet 5.5, with no Terra figure. HLE-Diamond is 58.8%, against 53.3% and 47.3%[4].

Robby Stein, VP of Product for Google Search, calls this the biggest model upgrade to Search yet. Hydrogen is on Search globally, for all users[3]. Google says Search can follow a question with several parts, hard-to-understand typos or a highly specific request, and answer each part. Query fan-out runs simultaneous queries across subtopics, which Google says returns a wider set of links. Answers are meant to include more images, comparison tables and custom tools, with more nuance in every supported AI Mode language[3].

Against 3.8 Flash, Google says Hydrogen cuts overreach in agentic tasks by up to 81% and under-service by up to 79%, and cuts misaligned actions and task shortcutting on tool-use and coding evaluations by 12 times. On impossible or blocked tasks it is up to two times less likely to bypass a permission boundary, with a 96.7% honesty rate on an internal evaluation. Chain-of-thought legibility held at 99.9% across more than 200,000 audited turns, with no deceptive reasoning edits. Google also says Hydrogen is more than 2.3 times more resilient on the Gray Swan indirect prompt injection benchmark, and that false refusals on HealthBench Pro stay under 0.4%[1].

The model card says unjustified refusals are higher than on 3.8 Flash, by 5.57 percentage points on a check where lower is better, particularly for multimodal prompts. A full safety and security report is listed as coming soon. Hydrogen showed no material increase in general capability over Argon, so a full suite of critical capability evaluations was not required. It does not reach the CBRN or cybersecurity critical capability levels, though it reaches the alert thresholds for both[4].

Hydrogen is the fast tier under Carbon and Argon, at a flat US$1.50 and US$7.50 across a 1 million-token window, with Orchestrator as Google's way to mix the fast model and the frontier one in a single session[1]. Free users of the Gemini app get Hydrogen when auto model selection chooses it[1]. The scores above are Google's own, and on Terminal-Bench 4.1 the model card lists Claude Sonnet 5.5 higher[4].