Google has released Gemini 4 Hydrogen, a fast model for coding and agentic work. Gemini 4 Hydrogen is a proprietary model from Google DeepMind built for everyday agent loops. It sits below the frontier models Gemini 4 Carbon and Gemini 4 Argon. Google calls it its “daily driver for agentic workflows.”
TL;DR
- Size: 1,048,576-token unified window for input, thinking and output.
- Runs on: Google-hosted API only. No weights, no quantized variants, no self-hosting. Serving hardware not disclosed.
- Performance: Scores 51 on the Artificial Analysis Intelligence Index at high, above Argon’s 50.
- Best: 57% on Terminal-Bench 4.1, versus 18.2% for Gemini 3.8 Flash.
- Worst: Ranks only #5 on Agent Arena. Costs 2x Gemini 3.8 Flash’s current introductory price.
- Bottom line: Best: frontier-class agent scores at $1.50/$7.50 per 1M tokens. Worst: closed, and synchronous output caps at 65,536 tokens.
What is Gemini 4 Hydrogen?
Hydrogen is a closed-weights model served through the Gemini API and Google products. The model card says it is adapted from Gemini 4 Carbon. Google does not disclose parameter count, architecture or training hardware. The knowledge cutoff is March 2026.
Input covers text, images, audio, video, PDFs, tables and charts, per the What’s New guide. Output is text. The 1M window is shared by input, thinking and output tokens.
How do the 4 thinking levels work?
Developers set thinking_level inside generation_config. Values are low, medium, high and a new xhigh. The default is high, not medium as on 3.8 Flash. A minimal setting now returns an error. The model card warns that higher levels use more tokens.
How does Hydrogen work with Carbon and Argon?
The key design idea is role-splitting across models. In Google Antigravity, an Orchestrator option switches models inside 1 task. Carbon explores, resolves ambiguity and plans. Hydrogen edits files, updates packages and runs tests. When Hydrogen hits friction, it hands control back to Carbon.
On Terminal-Bench 4.1, this pairing reached near-Carbon task success rates. It used 24% less quota and finished 19% faster than Carbon alone. The Antigravity SDK exposes the same split for custom multi-agent systems.
How does Gemini 4 Hydrogen perform on benchmarks?
All figures are self-reported by Google unless noted.
Speed is the headline claim. Hydrogen finishes tasks 2.5x faster than 3.8 Flash at the default high level. In partner tests, it ran 2x to 2.8x faster with 12% to 23% fewer tokens.
Google also reports more than 5x fewer fabricated claims than 3.8 Flash. It scored 96.7% on an internal agent honesty evaluation. It was 2.3x more resilient to indirect prompt injection on Gray Swan’s benchmark.
‘;});document.getElementById(‘g4h-bars’).innerHTML=h;document.getElementById(‘g4h-n3’).textContent=N[i];setTimeout(function(){q(‘.bars .f’).forEach(function(f){f.style.width=f.dataset.w+’%’});},40);post();}
q(‘.sw button’).forEach(function(b){b.onclick=function(){q(‘.sw button’).forEach(function(x){x.classList.remove(‘on’)});b.classList.add(‘on’);drawBars(+b.dataset.s);};});
function f(n){return ‘$’+n.toLocaleString(‘en-US’,{maximumFractionDigits:0});}
function calc(){var i=+document.getElementById(‘g4h-in’).value,o=+document.getElementById(‘g4h-out’).value;document.getElementById(‘g4h-iv’).textContent=i;document.getElementById(‘g4h-ov’).textContent=o;document.getElementById(‘g4h-c1’).textContent=f(i*1.5+o*7.5);document.getElementById(‘g4h-c2’).textContent=f(i*0.75+o*3.75);document.getElementById(‘g4h-c3’).textContent=f(i*2+o*10);}
document.getElementById(‘g4h-in’).oninput=calc;document.getElementById(‘g4h-out’).oninput=calc;calc();
window.addEventListener(‘load’,post);window.addEventListener(‘resize’,post);setTimeout(post,300);
})();
