Google Releases Gemini 4 Hydrogen, a Fast Model for Agent Workflows and Everyday Coding

Google introduced Gemini 4 Hydrogen on 12 November 2026, describing it as a model built for rapid execution across diverse agentic workflows, meaning multi-step tasks in which a model plans, calls tools and carries work through to completion, as well as everyday coding. The company's announcement places it in the Gemini 4 family as the fast execution model. Gemini 4 Carbon and Gemini 4 Argon are offered as frontier models for deep, long-horizon reasoning.

Access is broad. Hydrogen is available to all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. Paid Google AI Pro and Ultra subscribers can also use it in Google Antigravity, Google's agentic development environment.

The model card, published by Google DeepMind in November 2026, states that Hydrogen is adapted from Gemini 4 Carbon. It accepts text, image, audio and video files as inputs. Its context window, the total amount of text it can hold in one session, is 1 million tokens, a token being a small chunk of text roughly equal to a word or part of one. That window is shared across input and output. Default synchronous output is capped at 64K tokens, and background long decode can extend output to the full 1M. Pricing is flat at $1.50 per million input tokens and $7.50 per million output tokens, according to the blog post.

Google also introduced Orchestrator in Google Antigravity. Orchestrator dynamically switches between models such as Carbon and Hydrogen within a single session. On Terminal-Bench 4.0, Google reports that Orchestrator reached task success rates close to Carbon alone, while consuming 24% less quota on average and completing tasks 19% faster.

For teams that route work by cost and latency, latency being the delay before a response arrives, this is the mechanism that matters most. It makes the frontier-versus-fast split a per-step decision rather than a per-project one.

On speed, Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default setting. On Terminal-Bench 4.1, the blog post gives Hydrogen 57% against 18.2% for Gemini 3.8 Flash. Google also reports that Hydrogen cut fabricated claims by more than five times relative to Gemini 3.8 Flash, and that it outperforms similar-tier models on SWE-sweep, a benchmark for finding and repairing bugs without breaking existing code.

On the Artificial Analysis Intelligence Index, Hydrogen scored 51 at the default high thinking level, one point ahead of Gemini 4 Argon at 50. At its highest xhigh thinking setting, Hydrogen reached 54. Thinking level is a meaningful variable. The gap between default and maximum effort is larger than the gap between Hydrogen and Argon, so cost and latency budgets will shift depending on which setting teams choose.

The model card places Hydrogen against Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra across a broader set of evaluations. Hydrogen leads on several. It scored 13.5% on SWE-sweep, against 3.1% for Claude Sonnet 5.5 and 5.8% for GPT-6.1 Terra. On Agents' Last Exam v1.1 it scored 71.6%, against 60.4% and 69.8% respectively. On AutomationBench-AA it scored 76.1%, against 71.8% and 60.2%. On Vending-Bench 2 it reached a bank balance of $14,549, compared with $11,165 and $9,107. It also led on the MKA Climb WSR accuracy measure, at 53.4% against 52.0% and 43.8%, and on Chartography, at 66.6% against 61.6% and 63.2%.

Hydrogen trails on others. On Terminal-Bench 4.1 cheating-adjusted, it scored 57.0%, against 63.9% for Claude Sonnet 5.5. On the OSWorld-2.1 offline subset, it scored 77.6%, against 83.9% for Claude Sonnet 5.5 and 74.2% for GPT-6.1 Terra. On Terminal-Bench-Science 0.1 it scored 51.4%, against 59.9% and 58.3% respectively.

In my view, the more useful reading of this release is architectural rather than comparative. A fast model with a 1M-token window and a low per-token price is a practical default for high-volume agent loops, where the number of calls matters more than the depth of any single one. Orchestrator then provides the escalation path to Carbon when a task needs it. Whether the quota and speed savings hold outside Google's own Terminal-Bench 4.0 runs is a question for teams to test against their own workloads.

The broader context is that agentic coding has moved from demonstration to metered production use, where throughput, tool-call reliability and cost per completed task carry as much weight as leaderboard position. Hydrogen's published numbers speak to those constraints directly. The fabricated-claims reduction matters for the same reason. In an agent loop, a confident error at step three carries into every later step. What the release enables is a cheaper, faster default for routine coding and knowledge work, with a frontier model reserved for the hard cases, and that division of labour is now something a single session can manage automatically.