Sign in Subscribe
news

Gemini 4 Hydrogen

Sebastien Dubois

Sebastien Dubois

— 4 min read

Canonical version: Gemini 4 Hydrogen.

Gemini 4 Hydrogen is Google DeepMind's third Gemini 4 model, released on 12 November 2026, after Gemini 4 Argon and Gemini 4 Carbon. Argon and Carbon are the big, slow, expensive frontier models. Hydrogen is the fast one: Google calls it "our daily driver for agentic workflows", built for everyday coding and agent work. In practice, it's the Gemini 4 replacement for 3.8 Flash.

Logan Kilpatrick (Gemini API lead) summed up the goal: "an extremely capable and fast model built to bring Gemini 4 to everyone". Demis Hassabis added: "Incredible speed and intelligence for everyday coding, agentic work and all-round use."

That "everyone" part is what makes it matter. Argon and Carbon sit behind paid tiers and quotas. Hydrogen is in the free Gemini app, in Google Search, and in every tier of the API. For most people, Hydrogen IS what "Gemini 4" will mean.

Key facts

  • Pricing: $1.50 per million input tokens, $7.50 per million output tokens, flat across the whole context window (no surcharge for long prompts). That's the price of non-discounted 3.8 Flash
  • Context: 1M input tokens, up to 1M output tokens
  • Modalities: text, PDFs, spreadsheets, charts, audio and recorded video in; text and structured deliverables (e.g., cited reports) out. No image generation
  • Thinking levels: low, medium, high (default) and a new xhigh
  • Speed: about 2.5x faster than 3.8 Flash at the "high" setting
  • Availability: Gemini API and Google AI Studio (all tiers), Android Studio, Gemini for business, the Gemini app (model dropdown; free users get it through automatic model selection), Google Search (all users), Gemini in Google Sheets (AI Pro and Ultra first) and Google Antigravity (paid AI Pro and Ultra)

Benchmarks (Google's numbers)

Benchmark Gemini 4 Hydrogen Comparison
Artificial Analysis Intelligence Index 51 (high), 54 (xhigh) Argon 50, GPT-6.1 Sol 55, Claude Sonnet 5.5 55, GPT-6.1 Terra 51
Terminal-Bench 4.1 57% 3.8 Flash 18.2%
HLE-Diamond 58.8% Leads its class
Vals Tax Agent Bench 52.3% #1
Vals Legal Research Bench 55.8% Tied #1
Agents' Last Exam v1.1 71.6%

So a "Flash-class" model now beats Argon, Google's frontier model from two months earlier, on the Artificial Analysis index. At xhigh, it's one point behind Sonnet 5.5 and GPT-6.1 Sol. Artificial Analysis estimates about $3 to run its index at xhigh, versus $6-7 for 3.8 Flash at medium-high, and reports a higher cache hit rate. But Hydrogen uses more output tokens per task, so the per-token price doesn't tell the whole story.

Google also claims more than 5x fewer fabricated claims than 3.8 Flash, 12x fewer misaligned actions or task shortcuts in tool-use and coding evals, and 2.3x more resilience on the Gray Swan indirect Prompt injection benchmark. These are Google's own numbers, so wait for independent tests.

Orchestrator: Carbon plans, Hydrogen executes

The interesting part for agent builders came with it. Google Antigravity got an Orchestrator mode that switches between models inside one session. Carbon does a read-only planning phase, Hydrogen takes over the edits, package updates and tests, and the work goes back to Carbon when it hits friction. On Terminal-Bench 4.0, Orchestrator got close to Carbon's task success with about 24% less quota, and finished about 19% faster on average. In parallel-agent setups, Carbon can coordinate while Hydrogen runs the worker and verifier agents.

This is Model routing built into the harness: an expensive model for thinking, a cheap one for doing. I like that pattern a lot. On Hacker News, one commenter reported about 60% savings with a similar setup (Opus plus a cheap model), so Google's 24% isn't the ceiling. Watch out, though: another commenter noticed that picking Orchestrator switches permissions to "turbo" (yolo mode) by default. Orchestrator needs Carbon, so it's Ultra-only for now.

Reception on Hacker News

Two threads (565 points for the launch, 500 for the Antigravity post) said roughly the same things:

  • Good model for the price. People praised its image-to-HTML work ("between sonnet and opus") and its multilingual support. One data analytics exam reported a perfect score, after only GPT-6 Astra, GPT-6.1 Sol and Fable 5.2, at about a third of their cost
  • The naming is a mess. Hydrogen, Argon, Carbon: the periodic-table order doesn't match the model order ("Hydrogen --> Argon --> Carbon!?"). Several people wished for noble gases, or simply OpenAI's Luna/Terra/Sol/Astra ladder
  • Benchmarks vs reality. Some suspect Benchmaxxing ("google tends to benchmaxx"), and small models often look better on leaderboards than in real use
  • Harness and product sprawl. Many said the model is fine but Google's harnesses and its stack of products (Gemini app, AI Studio, Antigravity, Antigravity IDE, the agy CLI) are confusing
  • Account bans. Some users got suspended for calling the Antigravity CLI from another tool (access was later restored). Getting banned from Gemini can mean losing your whole Google account, Gmail included. That risk is very real
  • Prices going up. Some commenters said Google's pricing page announces a 2x price increase for Google models from 1 January 2027. Others disputed that it applies to Hydrogen. I couldn't confirm it

My take

Hydrogen is the most important of the three Gemini 4 models, even if it's the smallest. Most people never pick a model: they use whatever the free app or Search gives them. Giving them a model close to Sonnet 5.5 for free is a much bigger jump in "perceived AI intelligence" than a one-point lead at the top of a leaderboard.

For my own work, the Orchestrator pattern matters more than the model. Plan with a big model, execute with a cheap one: that's where agent costs go down. I'd still keep my Google account far away from any agent setup that could break the terms of service.

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe