Google DeepMind launches Gemini 4 Hydrogen for agentic coding
A fast execution model priced at $1.50 per million input tokens, positioned to pair with Carbon and Argon rather than replace them.
What Google announced
Google DeepMind introduced Gemini 4 Hydrogen on 12 November 2026, with the announcement page carrying a last-updated date of 13 November 2026. Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, framed the release around speed and efficiency for everyday coding and agentic workflows. The model sits below Gemini 4 Carbon and Gemini 4 Argon, which the post describes as frontier models built for deep, long-horizon reasoning on hard problems. Hydrogen is the execution layer: writing code, running tests, verifying changes and keeping multi-step agent loops moving.
The full announcement says Hydrogen is available across all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and to paid Google AI Pro and Ultra subscribers in Google Antigravity.
Benchmarks, Orchestrator and the multi-model argument
The headline claim is that Hydrogen edges out Argon on the Artificial Analysis Intelligence Index at its default high thinking level, with 51 against 50, and reaches 54 at a new xhigh setting. It also ships alongside Orchestrator in Google Antigravity, which switches between models such as Carbon and Hydrogen inside a single session instead of running one model from start to finish. On Terminal-Bench 4.0, Orchestrator reported near-Carbon task success rates while using 24% less quota on average and finishing 19% faster. For parallel work, the post suggests Carbon coordinating while Hydrogen powers worker and verifier agents.
For teams still on Gemini 3.8 Flash, the comparison numbers are the ones worth reading. DeepMind says Hydrogen finishes tasks 2.5 times faster at its high default setting, scores 57% against 18.2% on Terminal-Bench 4.1, cuts fabricated claims by more than five times, and recovers cleanly when a tool call fails or a search returns nothing. Early partner testing cited in the post found 2x to 2.8x faster runs with 12% to 23% lower token usage in graph, SQL and data-heavy enterprise workflows.
Thinking levels, context and price
Hydrogen has four thinking levels: low, medium, high (the default) and xhigh. The point is to pay only for the reasoning depth a task needs. The context window is 1 million tokens on the input side, with up to 1 million output tokens, which the post ties to synthesising PDFs, spreadsheets, charts, audio and recorded video into cited reports. Pricing is flat across that window at $1.50 per million input tokens and $7.50 per million output tokens.
What it changes for people building apps
Two things stand out. First, cost and latency now scale in two separate dials: the thinking level and the choice of model. A builder can route routine edits and test loops to the cheaper model and reserve the frontier models for design decisions or stubborn bugs. Second, if Orchestrator-style routing becomes normal, the model becomes something the harness picks rather than something you pick once per project.
The post gives a handful of concrete examples, all of them Google-side demos. A project called Geminisland lets anyone add structures to a shared voxel island from a single prompt. With a new front end design skill in Google Antigravity, Hydrogen built an e-commerce storefront in one shot with a WebGL2 shader, scrollytelling, a split-slider and a live configurator. Hydrogen also turned Google Sheets into an animated pixel canvas using Python, native formulas and conditional formatting, and produced a multi-slide presentation from a YouTube keynote, then ran its own audit-and-fix loop on the slide layouts. In Sheets, the post reports financial models generating outputs over 30% faster than previous models.
On the consumer side, the post says Hydrogen follows instructions more closely across multi-turn conversations in the Gemini app and Search.
Limits and open questions
The benchmark figures are self-reported, including the Artificial Analysis and Terminal-Bench results, and the partner speed and token savings come from early testing rather than a published method. The price is clear, but rate limits, regional availability, and whether free-tier Gemini app users get Hydrogen are not stated. The published text also cuts off mid-sentence in the section on Gmail and other Google apps, so the consumer rollout details are incomplete. Nothing in the post says how Hydrogen performs on non-Google coding environments.
For builders working with AI
If you are assembling an app with an AI builder, model choice is becoming a routing decision rather than a one-time setting. Draftship's paid plans include a choice of AI model, and the docs on using the AI chat and choosing a model cover how that selection works. Draftship is not affiliated with Google, and nothing in the announcement says Draftship supports Gemini 4 Hydrogen.
Primary sources
- Introducing Gemini 4 Hydrogen · Google DeepMind Blog, Nov 12, 2026
- Gemini 4 Hydrogen in GitHub Copilot · GitHub Changelog, Nov 13, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.