Introduction: speed stopped being a side feature
I build WordPress and Elementor sites for a living, and most of my working day is now a conversation with an AI model. It reviews my PHP, writes JavaScript, drafts copy, and debugs the plugin conflict that only shows up on the live server. The model matters less than it used to — they are all good now — but speed matters more than ever, because I wait on forty answers a day, chained together, while a client waits on me.
So when Google launches a model whose entire pitch is speed, I pay attention. On November 12, 2026, Google DeepMind launched Gemini 4 Hydrogen for everyday coding and agentic work, claiming it finishes tasks 2.5 times faster than Gemini 3.8 Flash. It costs $1.50 per million input tokens, ships with four thinking levels, and outscores Gemini 4 Argon on one independent benchmark. That is the press release version. The 2.5x number comes from Google's own task set, and the first independent measurements tell a more modest story.
What Google actually shipped on November 12
Gemini 4 Hydrogen is the third named member of the Gemini 4 family, and Google is clear about where it sits. Carbon and Argon remain the frontier models for deep, long-horizon reasoning. Hydrogen is the rapid-execution member: the model you reach for when the job is doing things, not thinking about them. Google DeepMind SVP Koray Kavukcuoglu describes it as bringing "cutting-edge intelligence to everyone, with speed and efficiency for everyday coding and agentic work."
The model card is explicit about lineage: Hydrogen is adapted from Gemini 4 Carbon, the frontier model in the same family. That matters, because Hydrogen is not a small model dressed up as a fast one. It is a frontier-class model tuned for throughput, with four thinking levels — low, medium, high, and xhigh — and high as the default. At that default, Google claims 2.5x faster task completion than Flash, and reports 51 on the Artificial Analysis Intelligence Index — 54 at xhigh — against 50 for Argon.
Hydrogen is in the Gemini app's model picker as "Everyday help," free accounts reach it through Auto, and the rollout is staged. On the API side it is generally available now, with the model ID
gemini-4-hydrogen and a one-million-token context window. The model card states a March 2026 knowledge cutoff, and a Safety and Security Report is promised rather than published.The 2.5x number: what it measures, and what it doesn't
"2.5x faster" can mean three different things, and Google's claim is only one of them. The first is throughput: tokens produced per second. On the independent Artificial Analysis harness, Hydrogen produces 190 output tokens per second, against 112 for Gemini 3.8 Flash, 65 for Argon, and 72 for Carbon. That is a real, measured advantage — roughly 70% more output per second. It is not 2.5x, but you feel it in a long coding session.
The second meaning is time per completed task, and here the picture flips. Artificial Analysis finishes an Intelligence Index task with Hydrogen in 13.5 minutes, with Flash in 14.8, Carbon in 28.4, and Argon in 30.0. That is about a 10% difference against Flash, not 150%. The reason is verbosity: Hydrogen generated 304 million output tokens to complete the index, against 72.8 million for the median model and 126 million for Flash. It thinks out loud, at length, and that verbosity eats the raw speed advantage. First-token latency is essentially a tie: 16.7 seconds at default high, against 17.3 for Flash.
The third meaning is Google's own internal task set, which is where the 2.5x figure comes from. Google says unnamed partners saw runs 2x to 2.8x faster, with 12% to 23% fewer tokens, and Sheets financial-model generation more than 30% faster. If the 2.5x figure came from a specific internal task set, it is a claim about that task set. No one outside Google has reproduced the gap on a general-purpose harness yet. The 2.5x sentence will get repeated, so plan around the measured numbers instead of the headline.
The price tag: what fast actually costs
Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across its one-million-token context window. Gemini 3.8 Flash is still in the catalogue at $0.75 per million input and $3.75 per million output — and Google has not retired it. Hydrogen costs exactly twice as much as Flash per token. That is the trade stated plainly: double the price, for a model that is faster, more capable on agentic benchmarks, and five times less likely to fabricate claims, by Google's own measurement.
Artificial Analysis puts the cost of completing one Intelligence Index task at $3.16 at xhigh and $1.87 at the default high setting. But remember the verbosity problem: Hydrogen generates far more output tokens per task than the median model, and output tokens are the expensive ones at $7.50 per million. A model that is twice the price per token and generates more than twice the tokens per task is not a cheap model. It is a fast model, and fast is what you are paying for. A WordPress debugging session is tens of thousands of tokens, not millions, so one session costs fractions of a cent either way. The cost only bites at volume — an agent loop running all night, a bulk content pipeline — and that is where dropping the thinking level from high to medium or low is the cheapest lever you have.
The benchmarks, read honestly
Every number in Google's launch table is vendor-reported and unreproduced, and the comparison set is Claude Sonnet 5.5 and GPT-6.1 Terra — not the Flash model being replaced. On Terminal-Bench 4.1, a test of AI agents working in a terminal, Hydrogen scores 57%, against 18.2% for Gemini 3.8 Flash. That is a genuine, dramatic jump — more than triple — and it is the number that matters most for developers, because terminal work is what coding agents actually do. But the same table shows Claude Sonnet 5.5 at 63.9% — a huge step up from Flash, and still not the best agent in the world.
On the Intelligence Index it is 51 at default high and 54 at xhigh, against 50 for Argon — a narrow lead over its own sibling, not over the frontier of the field. Google also says Hydrogen produces more than five times fewer fabricated claims — its term for made-up statements — than 3.8 Flash, which matters more in practice than a benchmark point: an assistant that invents a function costs you more time than a slow one.
The most useful independent data point is not a benchmark at all. Theo, the developer behind t3.gg, ran the same code review brief through several models: Hydrogen took 4m 43s, Carbon 14m 08s, Opus 5.5 2m 53s, GPT-6.1 Sol 2m 17s. One task, one data point — Hydrogen roughly three times faster than its own frontier sibling, while the fastest rivals still beat it.
Hydrogen vs the rest of the Gemini 4 family
If you are choosing a model today, the family has a clear shape. Gemini 4 Carbon is the frontier model Hydrogen was adapted from: the deepest thinker, the slowest of the three on raw throughput at 72 output tokens per second, and the one you reach for when a task genuinely needs long-horizon reasoning. Argon, the other frontier model, scores 50 on the Intelligence Index against Hydrogen's 51. Flash is the outgoing workhorse: still available, still half the price of Hydrogen, and now clearly outclassed on agentic work, with a Terminal-Bench score of 18.2% against Hydrogen's 57%.
Hydrogen is the middle of the pyramid: fast enough to feel instant, smart enough to beat its own flagship sibling on a combined benchmark, cheap enough to run at volume. Google has split the old Flash tier in two: a cheap tier, still Flash, and a fast tier, now Hydrogen. If your work is routine — drafting, summarising, simple code — Flash remains the rational choice. If your work is agentic — tools, terminal commands, multi-step tasks — Hydrogen is the model that was built for it.
What this means for WordPress and Elementor builders
I run a WordPress and Elementor site, and I build them for other people, so the question I care about is not whether Hydrogen is 2.5x faster in a lab. It is whether my day gets faster, and whether my sites get better. First, the coding loop. Most of my AI-assisted development is short, iterative work: a PHP snippet, a CSS fix, a JavaScript interaction, a plugin conflict. Hydrogen's throughput — 190 tokens per second against Flash's 112 — is the difference between watching code appear and waiting for it, and the Terminal-Bench jump from 18.2% to 57% matters because my coding agents live in the terminal. A model three times better at that loop is a better pair programmer, not just a faster one.
Second, the agent pattern. The most interesting idea in the launch is Hydrogen orchestrated with Carbon and Argon: the fast model doing the bulk work, the frontier model checking it, at 24% less quota and 19% faster than Carbon alone. That maps onto how I already work — a fast model drafts, a stronger model reviews — and any WordPress builder can copy it with the tools that already exist. Third, the content pipeline. Five times fewer fabricated claims means I can trust it further with first drafts — product descriptions, documentation, changelog copy. I still would not publish raw output — AI drafts, a human ships — but the review burden gets lighter. Fourth, cost discipline: the thinking level is the control — low for routine drafting, high for debugging, xhigh only when the task justifies the spend and the 37.8-second wait for the first token.
What I would do with it this week
I am pointing my routine coding-assistant traffic at Hydrogen at the default high thinking level, and keeping Carbon for the genuinely hard problems — architecture decisions, security reviews, the kind of task where being wrong is expensive. I am not migrating anything that runs on Flash and works, because Flash is still there, still half the price, and still fine for routine work. Before moving volume workloads I want independent re-runs of the fabrication and Terminal-Bench claims, and to see how the staged rollout lands.
The honest verdict: a real step forward, with throughput, agentic coding, and fewer fabrications all pointing the right way. The 2.5x headline is a claim about Google's own tasks, not a promise about yours — the independent harness puts the real-world gap over Flash closer to 10% on task time. Adopt it with your eyes open.
Conclusion: fast is a feature now, not a footnote
For years, model launches were about being smarter. Hydrogen is a launch about being faster, and that is a sign of where the industry is. The models are good enough now that the bottleneck is not intelligence but throughput — how quickly an agent can work through a job, how many jobs it can run in a day, how little you wait between asking and getting. Google has built a model for that world, adapted from its frontier model rather than cut down from it, priced for volume, and honest enough in its model card to state its limits.
I will use it daily from this week — high for coding, low for drafting, Carbon in reserve for the hard problems. If the independent numbers hold up over the next month, Hydrogen becomes my default workhorse the way Flash was before it. And if they don't, I will say so — the only benchmark that finally matters is the one measured in your own workflow, with your own clock running.
FAQ
What is Gemini 4 Hydrogen?
Gemini 4 Hydrogen is Google's speed-first AI model, launched on November 12, 2026. It is adapted from the frontier Gemini 4 Carbon model and built for everyday coding and agentic work, with four thinking levels and a one-million-token context window.
How much faster is Gemini 4 Hydrogen than Gemini 3.8 Flash?
Google claims it finishes tasks 2.5x faster than Gemini 3.8 Flash at the default high thinking level. Independent measurements from Artificial Analysis show 190 output tokens per second against 112 for Flash, but only about 10% faster on time per completed task, because Hydrogen generates far more output tokens per task.
How much does Gemini 4 Hydrogen cost?
The API price is $1.50 per million input tokens and $7.50 per million output tokens, flat across the one-million-token context window. That is exactly twice the price of Gemini 3.8 Flash, which remains available at $0.75 and $3.75 per million tokens.
What are the four thinking levels in Gemini 4 Hydrogen?
Low, medium, high, and xhigh, with high as the default. Higher thinking levels score better on benchmarks — 54 on the Intelligence Index at xhigh against 51 at high — but cost more tokens and take longer to produce the first answer.
Is Gemini 4 Hydrogen better than Gemini 4 Carbon and Argon?
On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at default high and 54 at xhigh, against 50 for Argon. But Carbon and Argon remain Google's frontier models for deep, long-horizon reasoning, and Hydrogen is positioned as the rapid-execution model, not the smartest one.
How can WordPress and Elementor users use Gemini 4 Hydrogen?
It is available through the Gemini API with the model ID
gemini-4-hydrogen, and in the Gemini app's model picker as "Everyday help," with free accounts reaching it through Auto. Developers can use it inside coding assistants and AI agents for PHP, JavaScript, debugging, and content drafting, and can pair it with a frontier model for review.Sources: Google DeepMind launch announcement and Keyword blog post (Koray Kavukcuoglu, November 12, 2026); Gemini 4 Hydrogen model card; Artificial Analysis Intelligence Index and throughput measurements; iGeeksBlog launch coverage; OrcaRouter analysis of the 2.5x claim; Aireiter Hydrogen vs Argon comparison; TechCityAuthority launch report; BusinessOutreach feature and pricing coverage; Theo (t3.gg) hands-on code review timing.
