Gemini 4 Hydrogen: Google's Fast, Low-Cost Model for Agentic Workflows

First came Gemini 4 Argon, the big general-purpose model it recommended for daily coding. Then came Gemini 4 Carbon, the frontier model for the hardest problems.

Now, there's Gemini 4 Hydrogen.

The name is the tell: hydrogen is the lightest element there is, and this is the light, fast, low-cost model in the group. Hydrogen is built to get things done quickly and cheaply, for the everyday coding and agentic work that doesn't need a frontier model's full reasoning budget.

The headline claim is that it matches or beats the much larger Argon on intelligence while finishing tasks around 2.5x faster than the Gemini 3.8 Flash model it replaces.

What Is Gemini 4 Hydrogen?

Gemini 4 Hydrogen is a distilled, speed-optimized model derived from Gemini 4 Carbon, Google's frontier model announced two days earlier. Where Carbon and Argon are designed for hard, long-horizon reasoning, Hydrogen is designed for execution: writing and editing code, running tests, verifying changes, carrying out multi-step research, calling tools, and analyzing documents.

It's a natively multimodal model that accepts text, images, audio, and video, with a knowledge cutoff of March 2026. Like its siblings, it supports a 1-million-token context window, and Google is pricing that entire window at a flat rate rather than charging more for long prompts.

Where It Fits in the Gemini 4 Family

The Gemini 4 lineup now has three tiers:

The most interesting design choice is that Google isn't positioning Hydrogen purely as a standalone model. In Google Antigravity, the company's agentic coding environment, a new Orchestrator mode switches between Carbon and Hydrogen within a single session. Carbon does the planning in a read-only phase, Hydrogen does the edits and runs the tests, and control can hand back to Carbon if execution stalls. Google reports that this combination reaches near-Carbon task success on Terminal-Bench while using roughly 24% less quota and finishing about 19% faster.

That planner-executor split mirrors a pattern that's become common in production agent systems: use an expensive model to decide what to do, and a cheap, fast model to do it. We talked about this idea in our Muse Ember article, also. 

Gemini 4 Hydrogen Benchmarks: How Does It Perform?

Google's own reported numbers are strong for a model in this class. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default "high" thinking level and 54 at the new "xhigh" level. Argon scores 50, which is the basis for Google's claim that the smaller model has overtaken the larger one.

For context, Carbon scores 59, and Claude Sonnet 5.5 and GPT-6.1 Sol both score 55 at their maximum settings.

Other headline results Google published:

• Terminal-Bench 4.1: 57.0%, up from 18.2% for Gemini 3.8 Flash
• HLE-Diamond: 58.8%, which Google says leads its class
• Agents' Last Exam v1.1: 71.6% (Carbon scores 77.4%, Argon 69.0%)
• AutomationBench: 76.1%
• OSWorld-2.1 (offline): 77.6%
• Vals Tax Agent Bench: 52.3%, ranked first overall
• Vals Legal Research Bench: 55.8%, tied for first
• Multimodal Knowledge Agent: 53.4%
• Agent Arena: fifth place

Independent testing from Artificial Analysis broadly confirms the picture while adding some caveats. Their Terminal-Bench 4.0 run puts Hydrogen at 54%, behind Claude Sonnet 5.5 at 64% and GPT-6.1 Sol at 56%, and Google's own model card notes that Sonnet 5.5 scores 63.9% on a cheating-adjusted Terminal-Bench 4.1 table versus Hydrogen's 57.0%.

In other words, Hydrogen is competitive on agentic coding, but it isn't the best model available for terminal-heavy work. Its clearest wins are in knowledge work, multimodal reasoning, and enterprise-style document tasks.

The hallucination story is also notable. Google says Hydrogen produces more than 5x fewer fabricated claims than Gemini 3.8 Flash, and Artificial Analysis measured an 8% hallucination rate on its AA-Omniscience benchmark.

Gemini 4 Hydrogen Speed and Thinking Levels

The appeal of Hydrogen is throughput. As I mentioned, it completes tasks about 2.5x faster than Gemini 3.8 Flash at the default setting. What I didn't yet mention: Early enterprise partners working on complex graph, SQL, and data-heavy workloads reported 2x to 2.8x speedups with 12% to 23% lower token usage.

Developers get four thinking levels to trade off quality, latency, and cost:

• Low for simple, high-volume tasks
• Medium for routine work where speed matters more than depth
• High (the default) for most coding and agentic tasks
• xhigh, a new tier for harder multi-step problems

Google's guidance is to leave most tasks on high and reserve xhigh for cases that need it, warning that xhigh can overthink simple prompts. Artificial Analysis measured the cost difference clearly: roughly $0.29 per task at low, $0.80 at medium, $1.87 at high, and $3.16 at xhigh, with the Intelligence Index dropping to 41 at medium and 27 at low.

A technical detail: input, thinking, and output all share the 1,048,576-token window, and synchronous responses are capped at 65,536 tokens per call. Longer outputs are possible through background execution or client-side continuation.

Gemini 4 Hydrogen Pricing

Gemini 4 Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens on the paid API, with cached input at $0.075 per million tokens via implicit caching. There's also a rate-limited free tier.

Artificial Analysis estimates that works out to roughly one-third the cost of Claude Sonnet 5.5 at its maximum setting, though about 2.5x the cost of GPT-6.1 Sol at max.

Combined with the flat pricing across the full context window, it's a meaningfully cheaper option for long-context agentic jobs than the frontier models.

Agentic Behavior and Safety

Because Hydrogen is meant to run autonomously in agent loops, Google has published more detail than usual on how it behaves when given tools. Compared with Gemini 3.8 Flash, the company reports:

• Up to 81% less agentic "overreach" (doing more than asked) and up to 79% less "under-service" (doing less than asked)
• A 12x reduction in misaligned actions and task shortcutting in tool-use and coding evaluations
• Up to 2x less likely to bypass permission boundaries when a task is blocked
• More than 2.3x greater resilience to indirect prompt injection on the Gray Swan IPI benchmark
• 96.7% on an internal agent honesty evaluation

The model card is candid about limitations too. Unjustified refusals rose by about 5.6 percentage points versus 3.8 Flash in automated comparisons, hallucinations remain possible, and higher thinking levels increase both latency and token usage. On safety thresholds, Google says Hydrogen shows no material capability increase over Argon and stays below its critical capability levels for biological information and cybersecurity, but sits at the alert thresholds in both domains, with fuller safety documentation to follow.

Where You Can Use It

Hydrogen is rolling out across Google's developer, enterprise, and consumer surfaces at once:

• Gemini API and Google AI Studio: available on all tiers, including the free tier
• Google Antigravity: available to paid Google AI Pro and Ultra subscribers; the Orchestrator mode is initially limited to Ultra
• Android Studio: paid tier
• Gemini for business (Gemini Enterprise): via the API, the model dropdown, Antigravity, and Gemini in Google Sheets for eligible subscriptions
• Gemini app: in the model picker for subscribers; free users get it through automatic model selection when it suits the prompt
• Google Search: rolling out to all users globally, with better handling of multi-part and typo-heavy queries, more visual responses, and broader query fan-out
• Gemini in Google Sheets: starting with Google AI Pro and Ultra subscribers

Trying Out Gemini 4 Hydrogen

Yes, the model is live today, and you don't need a paid plan to test it: pick gemini-4-hydrogen from the model list in Google AI Studio, or grab an API key there and call it from the Google Gen AI SDK with pip install --upgrade "google-genai>=2.34.0".

To see what "fast execution" actually looks like, we gave it one prompt in AI Studio at the default HIGH thinking level and didn't touch the result. On theme, the prompt asked for a simulation of hydrogen atoms bonding into H₂ molecules and breaking apart again. I thought this would be a fun test, given the name.

Build a single self-contained HTML file (inline CSS and JS, no external libraries)
that animates hydrogen atoms in a 2D box on a canvas.

Requirements:
- 60 hydrogen atoms move with random velocities and bounce off the walls.
- When two free atoms collide below a bonding energy threshold, they bond into an
  H2 molecule and move together as a pair, drawn as two circles joined by a line.
- A "temperature" slider controls average speed. At high temperature, collisions
  carry enough energy to break H2 molecules back into free atoms.
- Show a live counter of free atoms vs H2 molecules, and a small line chart of the
  H2 count over the last 30 seconds.
- Use a dark background, make it look polished, and keep it at 60 fps.

Return only the HTML.Powered by

Here's what came back, unedited, with the temperature swept from 180 K up to 2,400 K and back down. Cyan dots are free atoms; violet pairs are H₂ molecules: