# Google introduces Gemini 4 Hydrogen for agent workflows

> The model is rolling out to developers and consumer products with shared context limits.

By BIGCHANGE.AI Editorial

Published: 2026-11-12T22:56:38.336Z
Updated: 2026-11-12T22:56:38.336Z
Canonical: https://bigchange.ai/blog/gemini-4-hydrogen-agent-workflows

![Illustration of a developer working at a laptop.](https://bigchange.ai/api/media/file/developer-v3.png)
AI-generated illustration for BIG CHANGE.

Google introduced Gemini 4 Hydrogen on November 12 for coding and agent workflows, rolling it out across developer and consumer products.

## The big change

- Google positions Hydrogen, adapted from Carbon, for faster execution. Its availability differs by product, so access to one Google service does not establish access through another.
- Input and output share one context limit. Developers must leave room for thinking and tool calls as well as the final answer, rather than filling the entire allowance with documents or code.
- API spending depends on both the rate and the number of tokens used. A useful comparison therefore measures cost alongside successful work, rather than assuming that a lower unit price guarantees a cheaper result.

## Where it is available

The [announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) lists AI Studio, the Gemini API, enterprise products and Search. Free Gemini app users receive Hydrogen through automatic selection when appropriate; Antigravity requires Google AI Pro or Ultra.

## Planning a request

Google Cloud’s [guide](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-hydrogen) specifies a shared 1,048,576-token ceiling, including input, thinking, function calls and the final response. Tokens are the units used to account for a model’s input and output. More input leaves less space for generation.

A foreground segment produces up to 65,536 output tokens. Longer generation requires continuation through supported APIs; the guide identifies `v1beta1` for continuation tokens. This matters when an application expects one uninterrupted response: it needs to handle continuation, not treat the first segment as the finished answer.

Thinking levels are low, medium, high and extra-high, with high the default. The older `thinking_budget` and `MINIMAL` settings return errors. Migration therefore requires checking configuration as well as changing the model name.

## What the bill measures

The [Standard API tariff](https://ai.google.dev/gemini-api/docs/pricing) lists a free tier and paid rates of $1.50 per million input tokens and $7.50 per million output tokens, including thinking. For illustration, 900,000 input tokens cost $1.35 and 100,000 total output tokens cost $0.75: $2.10 before other charges. That output would require continuation under the guide’s segment limit.

This arithmetic is not a measured job cost. Tool charges and service tiers can change the bill, and reasoning tokens count even when they are not part of the final answer. Teams should compare the cost of a completed, checked task rather than just the price of its visible response.

## Reading the evidence

Google’s [card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/) reports 13.5% on SWE-sweep and 57.0% on cheating-adjusted Terminal-Bench 4.1. These are vendor results; BIG CHANGE could not read the detailed methodology. The card acknowledges hallucinations and delays; its full safety report is forthcoming.

The independent [SWE-sweep project](https://github.com/facebookresearch/swe-sweep) asks agents to find and fix bugs in existing repositories without hints about their type or location. This differs from supplying a precise bug report: finding the problem is itself part of the work. A repair also needs to survive checks, rather than merely produce plausible-looking code. The benchmark’s design does not independently validate Google’s run.

For a team deciding whether to adopt an agent, a useful trial starts with representative work and explicit permissions. Record failed attempts and recovery as well as successful changes, and have a person review the result. A tool finishing its command is not the same thing as a correct repair.

## Sources

- [Google Hydrogen announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/) — Launch and product rollout.
- [Google DeepMind model card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/) — Model description, results and limitations.
- [Google Cloud developer guide](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-4-hydrogen) — Token limits and thinking settings.
- [Gemini API pricing](https://ai.google.dev/gemini-api/docs/pricing) — Standard rates and free tier.
- [SWE-sweep](https://github.com/facebookresearch/swe-sweep) — Independent benchmark project.