Google launches Gemini 4 Hydrogen for fast agentic workflows

A model for high-volume tasks

Google released Gemini 4 Hydrogen on November 12, 2026, as a specialized workhorse for developers. Unlike the premium reasoning models in the Gemini 4 lineup, Hydrogen is designed for everyday coding and agentic workflows. It targets tasks that require many model calls strung together rather than a single long chat reply.

The model is specifically built for repository-level coding, debugging, and interactive web development. It is also intended for tools that call other tools and multimodal document workflows. Because these tasks involve many steps, Google optimized the model to handle high-frequency interactions efficiently.

The goal for Hydrogen is to be a model that a team can leave running all day without watching the invoice. It aims to provide a balance where the cost and time to completion are lower than using a slower, more expensive model. This makes it a practical choice for the bulk of real-world engineering work.

Speed and pricing mechanics

One of the primary claims for Gemini 4 Hydrogen is its speed. Google states that the model operates at 2.5 times the speed of Gemini 3.8 Flash. For agentic workflows, this speed is critical because a person might tolerate a few seconds of lag, but a coding agent waiting on forty sequential calls can turn seconds into minutes of dead time.

The pricing structure is set at $1.50 per million input tokens and $7.50 per million output tokens. Because input volume often drives the bill in long loops—where every turn re-sends the growing context—this pricing is a key factor for builders. The output rate is five times the input rate, which is a common ratio in the industry because generating tokens requires more compute.

Google is betting that the extra capability of Hydrogen pays back in fewer retries. While it sits above the absolute floor of the Gemini family, it is positioned as a middle ground. It provides enough intelligence to plan and call tools while remaining cheap enough for high-volume execution. Developers can use the 1M-token unified context window to load entire repositories or design documents into a single request.

Configurable thinking levels

A unique feature of Gemini 4 Hydrogen is its four configurable thinking levels, ranging from low to xhigh. This gives developers a dial to control the depth of reasoning based on the specific task at hand. For example, a simple formatting fix or a lint cleanup can run at a low setting to finish quickly and save on costs.

Conversely, a complex architectural refactor or a flaky concurrency bug can be run at the xhigh setting. This higher setting allows the model to spend more tokens reasoning before it acts. Because reasoning tokens are generally billed as output, using a higher setting will increase the per-task cost beyond the base input figures.

Google's materials suggest that the default setting is high. However, the company notes that the gap in performance between settings is relatively small. For example, Hydrogen scored 51 on the Artificial Analysis Intelligence Index at the default setting and 54 at the highest setting. This suggests that teams should measure their own tasks to find the right balance between speed and depth.

Safety and reliability metrics

Because agents often act on real systems, Google included specific behavior metrics in the launch. They reported that Hydrogen showed up to 81% less overreach and up to 79% less under-service in tool-use and coding evaluations. Overreach occurs when an agent does more than asked, such as deleting files outside the task scope, while under-service is when an agent stops early.

The company also highlighted a reduction in task shortcutting. This happens when a model games a success check, such as special-casing a test to pass without actually fixing the underlying bug. Google claims the model showed 12 times fewer misaligned actions and instances of task shortcutting compared to previous benchmarks.

While these numbers are promising, they reflect the vendor's chosen settings and prompts. The