Gemini 4 Hydrogen bets that agents want speed first
Google released Gemini 4 Hydrogen, a model built for rapid execution across agentic workflows, with four thinking levels and lower prices than its predecessor tier, plus an Orchestrator that switches models inside a session. It is also the biggest model upgrade to Google Search to date.
- Google DeepMind introduced Gemini 4 Hydrogen for rapid execution across agentic workflows, available to everyone starting November 12, 2026. The model offers four thinking levels: low, medium, high as the default, and xhigh. Self-reported m:…7670 m:…2541
- Gemini 4 Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens, the company says. Self-reported m:…7670
- The company says Gemini 4 Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default setting and cuts fabricated claims by more than 5x compared with Gemini 3.8 Flash. Self-reported m:…7670
- Google also introduced Orchestrator in Google Antigravity, which dynamically switches between models within a session. The company says Orchestrator consumes 24% less quota and completes tasks 19% faster on average compared with using a single model. Self-reported m:…7670
- Gemini 4 Hydrogen is now available globally in Google Search for all users, and the company describes it as the biggest model upgrade to Google Search to date. Self-reported m:…2539
Speed is the product
Google DeepMind introduced Gemini 4 Hydrogen for rapid execution across agentic workflows, and the model is available to everyone starting November 12, 2026. The pitch is not bigger reasoning or a wider context window. It is speed: a model built to move fast through the loops of coding, debugging and tool use that make up agentic work, where every extra second per step multiplies across hundreds of steps. m:…7670 m:…2541
The company says Gemini 4 Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default setting. Hydrogen offers four thinking levels, low, medium, high and xhigh, with high as the default, so developers can trade speed against depth per task instead of per model. The model handles repository coding and debugging, interactive web development and multimodal workflows, according to Google, which positions it as a general workhorse rather than a specialist. m:…7670 m:…2541
Gemini 4 Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens. The pricing matters because agentic work burns tokens in long loops, and a cheaper fast model changes what teams can afford to automate. A model that is 2.5 times faster and cheaper per token makes it practical to run agents on jobs that were previously too slow or too costly to hand over. m:…7670
Fewer invented claims, measured the company's way
Google says Gemini 4 Hydrogen cuts fabricated claims by more than 5x compared with Gemini 3.8 Flash, achieves a 96.7% honesty rate on internal agent honesty evaluations, and reduces overreach in agentic tasks by up to 81% compared with 3.8 Flash. These are the company's own measurements, and they point at the failure modes that matter most for agents: inventing facts, claiming work it did not do, and acting beyond what it was asked to do. m:…7670 m:…2541
The model ranks #5 on Agent Arena, and Google says it is 2.3x more resilient on the Gray Swan IPI benchmark against indirect prompt injection attacks. The injection result is the more interesting of the two for anyone shipping agents that read web pages, files and tool outputs, since those are the channels through which hostile instructions reach a model. A fifth-place arena ranking is a leaderboard position; resistance to injected instructions is a property that decides whether an agent can be trusted with real tools. m:…7670
The orchestrator switches models mid-task
Google introduced Orchestrator in Google Antigravity to dynamically switch between models within a session, pairing Hydrogen with Gemini 4 Carbon and other models in the family. The company says Orchestrator consumed 24% less quota on Terminal-Bench 4.0 and completed tasks 19% faster on Terminal-Bench 4.0, and 24% less quota with 19% faster completion on average compared with using a single model. The idea is that no single model is best for every step of a task, so the system routes each step to the model suited to it. m:…7670 m:…2541
Hydrogen is available in Google AI Studio, and it can run alongside Gemini 4 Carbon. That availability is the practical test: developers can try the speed and honesty claims on their own agent loops rather than taking the company's benchmarks at face value. m:…2541
Search gets the same model
Gemini 4 Hydrogen is now available globally in Google Search for all users, and Google describes it as the biggest model upgrade to Google Search to date. Advanced reasoning supercharges the query fan-out technique in Search, the company says, so Search conducts simultaneous queries across relevant subtopics and can now handle questions with multiple parts, questions with typos and highly specific requests. Search responses now include more images, structured responses, comparison tables and custom tools, and users can access a broader range of helpful links than before. m:…2539
The Search rollout is the distribution story behind the launch. A model that reaches every Search user on day one does not need developers to adopt it first; it changes what ordinary search can do, and the agentic features ride along. For developers, the question is whether the same model that answers search queries well also holds up inside long agent loops, and that answer will come from use, not from the launch numbers. m:…2539
What is still unresolved
- Whether the company's honesty and injection-resistance scores hold up on independent benchmarks
- How much Orchestrator's quota and speed savings vary outside Terminal-Bench 4.0
- Whether developers adopt the four thinking levels or leave the default on
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.