Key takeaways

  • Google says Gemini 4 Hydrogen can control a robot arm from a goal and a camera view by planning, writing and running control code, then adjusting it until the task is done.
  • A new open-source Robotics plugin for Google Antigravity handles arm setup, MuJoCo simulation, perception and app building; Google says a YAM arm was calibrated in about 10 minutes.
  • Google reports a score of 58 on Artificial Analysis Robotics 0.1 and an 84.8 average against 75.5 for Gemini Robotics ER 2 on 16 embodied reasoning tests; the API costs $1.50 per million input tokens and $7.50 per million output tokens.

Gemini 4 Hydrogen robotics is Google's pitch that its fast, cheaper Gemini 4 model can run a real robot arm: you give it a goal in plain language and a camera view, and it plans the task, writes the control code, runs it and corrects itself until the job is done. Google described this in a Google Developers Blog post on November 17, 2026, and launched a Robotics plugin for its Antigravity coding tool on the same day.

What's new is the price, not the skill. Google says the bigger Gemini 4 Argon could already drive an arm, and Hydrogen keeps that ability at a lower cost per token. Hydrogen itself came out on November 12 (our launch and pricing story).

What embodied and spatial reasoning mean

Spatial reasoning is understanding where things are in 3D: which cup is behind the bottle, how far the box is from the gripper, whether two camera views show the same object.

Embodied reasoning goes one step further. It is reasoning as if you had a body in that space: what to grab first, where to put the gripper, how a motion will play out over time, and whether the task actually succeeded. Google's tests for it include pointing at objects, reading gauges, breaking a task into steps and spotting when a step failed.

Until now Google sold these skills in a separate model, Gemini Robotics ER 2 (ER for embodied reasoning), a planner that hands the motor movements to a vision-language-action model, one whose outputs are motor commands (what a VLA model is). Google says those skills now live inside the main Gemini 4 series.

What Gemini 4 Hydrogen does for a robot

Hydrogen controls the arm the way a coding agent would, with a camera for eyes: it looks at the scene, plans the steps, writes and runs control code, then watches the result and fixes the code. Google's demos:

  • Setup in about 10 minutes. A developer plugs an I2RT YAM arm and a depth camera into a laptop, and the plugin installs software, finds the camera and calibrates the arm. Google says the same works for the low-cost SO-ARM101 and the industrial Franka FR3 Duo.
  • A snack-sorting conveyor in about an hour. From a sketch and photos of the scene, Hydrogen builds a simulation, tests grasps there, and the developer then runs the result on the real arm, tuning gripper force by chat.
  • A soap shop that packs its own orders with two YAM Pro arms.

Google says Hydrogen's lower cost makes it practical to run a robot application all day, but publishes no cost or uptime figure to back that.

The Robotics plugin for Google Antigravity

Google Antigravity is Google's agentic coding tool. The new plugin adds skills, packaged instructions and scripts the model can use, in four areas:

Skill area What it does, per Google
Robot bring-up Installs drivers, finds cameras, calibrates the arm and checks safe motion, starting from a USB cable
MuJoCo in Antigravity Builds and previews simulated scenes in MuJoCo, an open-source physics simulator, inside the editor
Perception and control Uses Gemini's spatial understanding to find objects, plan grasps and control the robot in a closed loop
Application building Generates the controller and simulation and moves the app from real scene to simulation and back to hardware

Google calls the simulation step "vibe sim": you change a simulated scene by chat and check that the arm can reach and grasp before touching hardware.

The plugin is open source on GitHub under Apache 2.0. Install it in Antigravity under Customizations (search for Robotics on the Build With Google shelf) or with agy plugin install. Simulation needs no robot, and skills that call Gemini's perception models need a GEMINI_API_KEY. The README warns that "a moving robot can injure people and damage equipment" and puts safe operation on the user.

Gemini 4 Hydrogen robotics benchmarks Google reports

Google cites two results, both in its own framing.

Artificial Analysis Robotics 0.1. In this test from Artificial Analysis, a model controls a real 6-axis arm through eight tabletop tasks, from pressing a button and pouring a drink to building the tallest tower and measuring a box. A human grades each end state out of 100. Google says Hydrogen scores 58, "similar to Argon at a lower cost per token," citing a post by Artificial Analysis on X. The benchmark page itself, published October 14, 2026, doesn't list Hydrogen yet. It lists these scores:

Model AA Robotics score
GPT-6 Astra 69.4
GPT-6.1 Sol 64.6
Gemini 4 Hydrogen (per Google) 58
Gemini 4 Argon 55.3
Claude Opus 5.5 48.8
Claude Fable 5.1 46.3

So by Google's own number, Hydrogen edges past Argon but still trails both OpenAI models in that test.

ERQA and other embodied reasoning tests. ERQA is a Google DeepMind set of 400 multiple-choice questions about images of real scenes. Google says Hydrogen beats ER 2 on spatial tasks such as matching points across camera views and reasoning about 3D paths. A chart in the post compares the two at high thinking across 16 tests, including ERQA, gauge reading, pointing and success detection: Hydrogen averages 84.8, ER 2 averages 75.5, and ER 2 stays ahead on steerable pointing. Google tells ER 2 users to try Hydrogen for these workloads.

How to access Gemini 4 Hydrogen for robotics, and what it costs

The model ID is gemini-4-hydrogen everywhere. Three ways in:

  • Google Antigravity with the Robotics plugin. Select Gemini 4 Hydrogen and install the plugin. Google's suggested first prompt: "Bring up my SO-101 arm and make it wave." Antigravity needs a paid Google AI Pro or Ultra subscription.
  • Gemini API. Use gemini-4-hydrogen through the Interactions API, with support in the google-genai SDKs from day one.
  • Google AI Studio. Try spatial prompts, live camera streams and structured tool calls in the browser.

API prices per million tokens, from Google's pricing page as of November 19, 2026:

Tier Input Output (incl. thinking) Cached input
Free Free Free Free
Standard $1.50 $7.50 $0.075
Batch $0.75 $3.75 $0.0375
Flex $0.75 $3.75 $0.0375
Priority $2.70 $13.50 $0.135

Unlike Gemini 4 Carbon, Argon and Neon, Hydrogen is on the free tier, where Google uses your content to improve its products. Limits from the model card: a 1 million-token context window, up to 64,000 output tokens per normal request, and a knowledge cutoff of March 2026.

What isn't known yet

  • Hydrogen's exact scores. The 58 comes from a post on X; the benchmark page doesn't show it yet or say how many trials each task gets. The chart prints no per-test numbers, so Hydrogen's ERQA score is unknown.
  • Speed. On the same benchmark Argon averaged 52 minutes per task, against about 20 seconds for a human. Google gives no time for Hydrogen, and Artificial Analysis writes that current models are far too slow and expensive for production robots.
  • Cost and success rates for the demos. Google publishes neither.
  • The future of ER 2. Google hasn't said whether the preview model will be retired.
  • Physical safety. The model card covers content and misuse, not moving hardware.

For a developer with a hobby arm, trying it costs little: Hydrogen is on the API free tier, the plugin is open source, and simulation needs no hardware. Google is collecting projects on the plugin's GitHub discussions page.

Frequently asked questions

What are the new features of Gemini 4 Hydrogen for robotics?

Hydrogen brings the embodied reasoning skills of Google's Gemini Robotics ER 2 models into the main Gemini 4 series and can drive a robot arm by writing and running control code. Google also launched a Robotics plugin for Antigravity that sets up arms, builds MuJoCo simulations and helps build robot apps.

How much does Gemini 4 Hydrogen cost?

On the Gemini API standard tier it costs $1.50 per million input tokens and $7.50 per million output tokens, as of November 19, 2026. Batch and Flex are half that, and the model is also on the free tier.

How does Gemini 4 Hydrogen compare with Gemini 4 Argon on robotics?

Google says Hydrogen scores 58 on Artificial Analysis Robotics 0.1, similar to Argon at a lower cost per token. The benchmark page lists Argon at 55.3.

Do I need a robot to use the Antigravity Robotics plugin?

No. The plugin's README says simulation-only use needs no hardware. Antigravity itself requires a paid Google AI Pro or Ultra subscription.

Sources

  1. Gemini 4 Hydrogen: a faster model for general purpose roboticsGoogle Developers Blog
  2. Robotics plugin for Google Antigravity (README)Google DeepMind
  3. Gemini 4 Hydrogen model cardGoogle DeepMind
  4. Gemini API pricingGoogle
  5. Introducing Gemini 4 HydrogenGoogle
  6. Gemini Robotics ER 2Google DeepMind
  7. ERQA benchmarkGoogle DeepMind
  8. AA-Robotics v0.1Artificial Analysis

Add ClaudeGPT Si as a preferred source on Google