Gemini 4 Hydrogen launches as Google targets fast coding and agents

Akhil Ponnada
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.
6 Min Read
Image: Google Antigravity

Google launched its Gemini 4 Hydrogen artificial intelligence model on 12 November 2026, beginning a worldwide rollout across search, developer tools and enterprise workspaces. The release replaces the company’s previous Flash tier, offering a lightweight system tuned for fast code generation, test verification and multi-step agentic workflows.

Splitting planning from execution

Hydrogen is adapted from Gemini 4 Carbon and takes on a specialised operational role in Google’s element-named Gemini 4 family, as reported by Unite.AI. While Carbon and Argon handle deep architectural decisions and extended reasoning, Hydrogen is designed for rapid execution.

The system accepts text, images, audio and video inputs. Its knowledge cutoff is March 2026, which Google acknowledges may leave it less current in fast-moving subject areas.

To divide work between models, Google introduced an Orchestrator option in Google Antigravity. In a single development session, Carbon can inspect dependencies and build a project plan in a read-only phase. Hydrogen then steps in to modify files, update packages and run software tests.

When automated execution hits an error, control shifts back to Carbon to inspect the environment and revise the plan. Unlike static routers that send an entire prompt to one model, Orchestrator changes the model as the task develops.

Google’s Antigravity engineering team said the pairing achieved near-Carbon success rates on the Terminal-Bench 4.1 coding benchmark while using 24 per cent less quota and finishing 19 per cent faster on average. Carbon can also manage parallel tasks, directing multiple Hydrogen agents acting as worker and verification units in isolated virtual environments.

Pricing and thinking controls

Google published developer rates in the Gemini API documentation alongside a rate-limited free tier. For paid access, queries cost $1.50 for every million input tokens, while generation runs at $7.50 for the same volume of output. Reused prompts billed through implicit caching cost $0.075 per million tokens.

Hydrogen operates with a context window of 1,048,576 tokens shared between input, reasoning and output. Standard synchronous calls deliver up to 65,536 tokens per request, while background runs or client continuations can fill the remainder of the one-million-token window.

Developers can set four thinking levels: low, medium, high and a new xhigh setting. Google sets high as the default for difficult technical work, while suggesting xhigh only for complex multistep problems where high fails.

Because input, reasoning and generated text share a single token budget, elevated reasoning limits the space left for prompts and final answers. Setting the thinking control to xhigh can also increase latency or cause the model to overthink simple instructions.

How does Hydrogen perform on benchmarks?

On the Artificial Analysis Intelligence Index, Hydrogen scored 51 points at its high thinking setting and 54 at xhigh. That compares with 50 points for Argon, meaning Hydrogen matches or exceeds its larger sibling on that metric while costing less, as Neowin reported when comparing the release with GPT-6.1 Terra.

Performance jumped sharply when measured against Gemini 3.8 Flash. Hydrogen completed tasks 2.5 times faster at the high setting, scored 57 per cent on Terminal-Bench 4.1 compared with 18.2 per cent for Flash, and reduced fabricated statements by more than fivefold.

Data published in the model card places those gains in perspective. In a cheating-adjusted Terminal-Bench 4.1 comparison, Anthropic’s Claude Sonnet 5.5 scored 63.9 per cent, outpacing Hydrogen’s 57.0 per cent score.

Search rollout and safety assessments

Google has started distributing Hydrogen across its product catalogue, including the Gemini mobile app, Gemini for business, Gemini in Sheets, Google Antigravity and the core search engine, as Google detailed on its Google Blog.

Hydrogen has replaced Flash in the Gemini model picker, according to XDA. Free users of the Gemini app will route through Hydrogen when Google’s automated systems select it for a prompt.

In search, Google said the model helps the engine “handle even more nuanced and complicated questions, quickly understanding what you need”. Search gains broader query fan-out for external links, more visual cards, and better recognition of typos and multi-part queries.

The upgrade lands as Google expands AI features across its services, following education disputes where Google Search in schools faces restrictions if districts reject AI features. In developer workflows, tool selection connects to the Connected Apps Platform that opened Gemini and Search to external services such as Gmail, Google Drive, Airtable and Dropbox.

Agentic safety tests showed reductions of up to 81 per cent in model overreach and 79 per cent in under-service compared with Gemini 3.8 Flash, alongside higher resistance to indirect prompt injection. However, automated testing recorded a 5.57-percentage-point rise in unjustified refusals.

Google’s frontier safety evaluation found no general capability jump over Argon. Hydrogen remained below the severe-risk thresholds for biological and cybersecurity hazards, though it reached the company’s alert levels in both categories. Google said further safety evaluations will follow.

Follow:
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.