Gemini 4 Hydrogen speeds up execution of common tasks
Google deploys Gemini 4 Hydrogen in its consumer and professional products. The model targets coding, research and daily assistance tasks, with Orchestrator to switch between models according to difficulty.
On November 12, 2026, Google began rolling out Gemini 4 Hydrogen, a model designed for fast execution of common tasks in coding, research, and conversational assistance. It launches alongside Orchestrator, an option in Google Antigravity that switches between Hydrogen and the group's most powerful models depending on session needs. The goal is simple: reduce cost and computation time without sacrificing too much quality. No need for a jackhammer to drive in a nail.
A lighter model, but not relegated to a secondary role
Hydrogen positions itself as the execution model of the Gemini 4 family, while Carbon and Argon handle the most complex tasks. Google presents it as suitable for everyday coding, checks, targeted refactorings, and agent workflows where speed matters as much as accuracy. The model is also supposed to better follow instructions, which interests both developers and general users who want a usable result on the first try.
In practice, Google highlights several gains measured on its internal and third-party benchmarks. On Terminal-Bench 4.1, Hydrogen scores 57%, compared to 18.2% for Gemini 3.8 Flash, and it reportedly completed some tasks 2.5 times faster at its default level. On the Artificial Analysis Intelligence Index, it scores 51 at the high level, then 54 at xhigh, slightly outperforming Argon at the standard level.
Orchestrator, or how to avoid running a large model full-time
Orchestrator is the other major announcement. In Google Antigravity, this option dynamically switches between Carbon and Hydrogen during the same task, instead of using a single model from start to finish. Google explains that Carbon can initiate the plan, explore dependencies, and resolve ambiguities, while Hydrogen then takes over to write, test, or correct. If needed, Carbon resumes control. The whole aims for a finer balance between quality, latency, and token consumption.
On Terminal-Bench 4.0, Orchestrator reportedly achieves a success rate close to Carbon, while consuming on average 24% less quota and finishing 19% faster. For teams juggling multiple agents, the calculation is simple: better to reserve the most costly model for justified steps rather than letting it run in a loop doing filler work.
A broad deployment, from code to Search
Google does not reserve Hydrogen for developers only. The model is announced in Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and Google Antigravity, with access varying by offer. For Google AI Pro and Ultra subscribers, Hydrogen becomes available in Antigravity; for businesses, it is accessed via Gemini for business and associated work interfaces.
On the Search side, Google aims to improve understanding of long, misspelled, or very specific queries. Responses should also become more visual and richer in relevant links. In Gemini, Hydrogen is presented as the model that helps draft emails, study guides, summaries, or documents with fewer back-and-forths.
Performance and pricing designed for daily use
The model is priced at $1.50 per million input tokens and $7.50 per million output tokens, with a native context window of 1 million input tokens and up to 1 million output tokens. Google describes it as the cheapest member of the Gemini 4 family, although the price remains close to that of more ambitious models according to pricing grids published by industry observers [to be verified].
On security, Google also emphasizes a reduction in overreach behaviors in agent tasks, with up to 81% less overreach and up to 79% less under-service compared to 3.8 Flash. The company also claims better resistance to indirect prompt injections and 99.9% preserved reasoning transparency over more than 200,000 audited sessions. These figures come from internal tests but indicate the direction: making AI more reliable when it acts, not just when it talks well.
By the numbers
- 57% — Gemini 4 Hydrogen score on Terminal-Bench 4.1, according to Google.
- 24% — average quota saved by Orchestrator on Terminal-Bench 4.0.
- 19% — average time saved by Orchestrator on Terminal-Bench 4.0.
- $1.50 — price per million input tokens for Hydrogen.
- 200,000 — sessions audited for reasoning transparency, according to Google.