Google Gemini 4 Hydrogen: Fast AI for Everyday Work

Updated: November 20266 min readBy the CreatorSprout team

Launch coverage based on Google’s official announcement and independent launch reporting. We have not tested Gemini 4 Hydrogen hands-on. Google’s own claims and independently reported details are kept separate below. No affiliate links in this article.

Quick answer

On November 12, 2026, Google DeepMind launched Gemini 4 Hydrogen, the fast execution model of the Gemini 4 family, and Google calls it its “daily driver for agentic workflows.” It costs $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1 million-token context window, and it is the first Gemini 4 model free users can actually get. At its default high thinking setting it scores 51 on the Artificial Analysis Intelligence Index, one point above Gemini 4 Argon, and 57% on Terminal-Bench 4.1, more than triple Gemini 3.8 Flash’s 18.2%. Unlike Gemini 4 Carbon, which launched to vetted cyber defenders only, Hydrogen is available to everyone from day one, in the Gemini API, AI Studio, the Gemini app, Google Search and Antigravity. The honest headline: this is the Gemini 4 model you can actually use today, and the one Google wants running your everyday coding agents.

Google’s Gemini 4 family now has three tiers, and Hydrogen completes the set. Argon is the general-purpose flagship, Carbon is the frontier specialist for cybersecurity and multi-day engineering work, and Hydrogen, announced November 12, is the fast one: built for the unglamorous half of agentic work — writing the change, running the test, reading the failure, trying again. Google says it is adapted from Carbon, which explains how a “fast” model can edge Argon on an independent intelligence index while costing less than both of its siblings.

What Gemini 4 Hydrogen is

Gemini 4 Hydrogen is Google DeepMind’s new model for rapid execution across everyday coding and agentic workflows, announced November 12, 2026, in a Google blog post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect. Google frames it as the execution engine rather than the strategist: the model that does the routine work in a multi-model setup, while a heavier model plans. Google also announced Antigravity Orchestrator alongside it, a feature that hands work between Carbon and Hydrogen automatically.

Hydrogen pairs a native 1 million-token input window with up to 1 million output tokens, which Google says lets it synthesize dense PDFs, spreadsheets, charts, audio and recorded video into cited reports and structured deliverables. The context window is 1,048,576 tokens in total, shared between input, reasoning and output. One practical detail: ordinary synchronous calls cap output at 65,536 tokens; the full 1 million output tokens are reachable only through background long-decode mode. The model’s knowledge cutoff is March 2026.

The 2.5x speed claim, checked

The number that will circulate from this launch is Google’s claim that Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash, the model it replaces as the fast tier, at its default high thinking setting. It is worth checking, because the independent evidence tells a more mixed story.

On raw output speed, Hydrogen really is fast: Artificial Analysis measured 190 output tokens per second, against 112 for Gemini 3.8 Flash, 65 for Argon and 72 for Carbon. But on time per completed task, the same independent harness measured 13.5 minutes for Hydrogen at xhigh versus 14.8 minutes for 3.8 Flash at high, roughly a 10% difference, not 150%. Part of the reason is that Hydrogen generates far more tokens to finish a task: 304 million output tokens to complete the index, against 126 million for 3.8 Flash. The honest read is that Hydrogen is the fastest Gemini 4 model at producing text, and it finishes benchmark tasks somewhat quicker than Flash, but Google’s 2.5x figure is a vendor claim measured Google’s way, not an independent result.

Benchmarks: what the numbers actually say

The independent headline is the Artificial Analysis Intelligence Index: Hydrogen scores 51 at its default high thinking level, one point above Argon at 50, and 54 at its highest xhigh setting. Carbon, the frontier model, scores 60. So Hydrogen is genuinely strong for a “fast” model, but the 51-versus-50 edge over Argon is thin, and it is measured at Hydrogen’s default setting.

Google’s own reported results, not independently reproduced:

Terminal-Bench 4.1: 57% versus 18.2% for Gemini 3.8 Flash. The benchmark tests completing tasks in a terminal, the core of agentic coding work. More than triple the old model’s result, and the single most dramatic number in the launch.
Terminal-Bench 4.0: 54% at high and xhigh. On this earlier version of the same benchmark, Hydrogen came close to Carbon’s success rate while using 24% less quota and finishing 19% faster, according to Google.
Fewer fabricated claims. Google says Hydrogen produces fabricated claims more than five times less often than 3.8 Flash, and reduces overreach in agentic tasks by up to 81% and under-service by up to 79%.
Knowledge work: 71.6% on Agents’ Last Exam v1.1, 58.8% on HLE-Diamond, #1 on Vals Tax Agent Bench at 52.3% and tied #1 on Vals Legal Research Bench at 55.8% (the Vals leaderboard itself shows 53.85% for the legal bench, a small discrepancy worth noting), and 53.4% on the Multimodal Knowledge Agent benchmark.
Agent Arena: Hydrogen ranked fifth at launch on the live leaderboard for multi-step agentic tasks.

For comparison, Anthropic’s Claude Fable 5.2, released November 9, scores 64.6 on the same Artificial Analysis index, so Hydrogen is not competing for the overall crown. Its pitch is value per dollar, not the top score: Artificial Analysis estimates $1.87 of inference to complete one index task at high thinking ($3.16 at xhigh), against $4.97 for Argon, $7.76 for Carbon and $7.77 for 3.8 Flash, which is expensive for its quality because of how many tokens it burns.

Four thinking levels

Hydrogen ships with four configurable thinking levels: low, medium, high (the default) and a new xhigh tier. The idea is a smooth trade-off between quality, latency and cost, so you only wait and pay for the reasoning depth a task actually needs. The Artificial Analysis scores show the range clearly: 27 at low, 41 at medium, 51 at high and 54 at xhigh, with estimated cost per index task rising from $0.29 at low to $3.16 at xhigh. For quick, routine agent steps, low or medium is the cheap setting; for the hard step in a workflow, xhigh is there. Google fixes the sampling parameters (temperature, top_p, top_k and candidate count) at documented defaults, so you cannot tune them yourself.

Pricing

API pricing: $1.50 per million input tokens and $7.50 per million output tokens, including reasoning tokens, flat across the entire 1 million-token context window. No long-context surcharge tier, and no introductory promotion, unlike Argon and Carbon.
Cached input: $0.075 per million tokens, implicit caching only with a 4,096-token minimum. Explicit context caching is not supported.
Free tier: unlike Argon and Carbon, Hydrogen has a free API tier with limits, and it is the first Gemini 4 model free users can get, including in the Gemini app and Google Search.
The catch on the old model: Gemini 3.8 Flash still costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, half of Hydrogen’s price. From January 1, 2027, Flash’s regular prices match Hydrogen’s exactly. So Hydrogen is the upgrade, but it is a pricier one than Flash until the end of the year.

Where you can use it

Now, globally: the Gemini API and Google AI Studio for developers, Gemini for business, the Gemini app, and Google Search, where Google calls it the biggest model upgrade yet. Paid Google AI Pro and Ultra subscribers also get it in Google Antigravity.
Model ID: gemini-4-hydrogen in the API.
First Gemini 4 model for free users: earlier Gemini 4 models reached free tiers slowly or not at all; Hydrogen is the first one rolling out to everyone, including free users, from launch day.
Watch out for fakes: any third-party site claiming to sell “Gemini 4 Hydrogen access” is not coming from Google. The model is in Google’s own products and API; there is nothing to buy from a reseller.

The limits Google’s docs add

Google’s announcement is enthusiastic; Google Cloud’s model documentation is blunter, and it adds four constraints the blog post does not lead with:

Short-term general availability. Hydrogen is a short-term GA release subject to a minimum 45-day retirement notice policy. Google can retire it with as little as 45 days’ warning, so do not build a product that cannot swap models.
No RAG Engine, no tuning, no Live API. The capability table lists all three as not supported.
Implicit caching only. Durable caching is enabled by default, but explicit context caching is not supported, and TTL-based cache control arrives only from January 12, 2027.
Fixed sampling, fixed output caps. Custom sampling parameters are ignored, and the default synchronous output cap is 65,536 tokens, with the remaining context reachable only through background long decode.

What it means for creators and developers

For most readers, Hydrogen matters more than Carbon did, precisely because you can use it. If you run coding agents, Hydrogen is the sensible default for the routine steps: writing code, running tests, verifying changes, doing research, carrying multi-step tasks to completion. The four thinking levels are the practical feature here, run cheap at low or medium for the easy steps, spend xhigh only on the step that decides the outcome.

The multi-model pattern Google is pushing is also worth understanding. In Antigravity, the new Orchestrator hands work between Carbon and Hydrogen: the heavy model plans, the fast model executes. You can copy that pattern in your own tooling with any agent framework: a frontier model for the plan, Hydrogen for the loop. And because Hydrogen is the first Gemini 4 model in Google Search and the Gemini app for free users, it is also the model most people will meet first, which makes it the one to know when clients or collaborators ask what changed.

The broader pattern continues: frontier releases are now arriving in weeks, not quarters. Argon reached general availability October 13, Carbon was announced October 28, OpenAI’s GPT-6.1 Terra landed October 26, Moonshot’s Kimi K3.1 Pro on October 23, Claude Fable 5.2 on November 9, and now Hydrogen on November 12. Benchmark leads change hands every few weeks. Pick a model on the workflow it actually improves for you, and re-test when the next release lands.

The honest caveats

We haven’t tested it. Everything here comes from Google’s announcement, its model card and launch reporting, not our own runs.
The 2.5x speed claim is Google’s, and the independent check is smaller. Artificial Analysis measured about 10% faster per task, not 150%, even though raw output speed (190 tokens/sec) is genuinely the fastest in the Gemini 4 family.
Most headline benchmarks are Google’s own. The Artificial Analysis index score is independent; the Terminal-Bench and knowledge-work figures are vendor-reported and not yet independently reproduced.
It can be retired with 45 days’ notice. Hydrogen is a short-term GA release. Budget for a model swap, not a permanent dependency.
Flash is still cheaper until New Year’s Day. If your workload runs fine on 3.8 Flash, Hydrogen costs double until December 31, 2026, when Flash’s prices rise to match.
No tuning, no RAG Engine, no Live API. If your product needs any of those, Hydrogen is not the model for it yet.

Verdict

Gemini 4 Hydrogen is the Gemini 4 launch that actually changes what you can build this week. Carbon took the headlines and Argon took the flagship slot, but Hydrogen is the one in your hands: fast, cheap per task, available everywhere including free tiers, and strong enough on independent benchmarks to edge Argon at its default setting. The 2.5x speed claim is overstated against independent measurement, and the 45-day retirement notice means you should treat it as a replaceable part, not a foundation. But as a daily driver for coding agents and everyday agentic work, at $1.50/$7.50 with a free tier, it is the most useful Gemini 4 model Google has shipped so far. Try it in AI Studio, run it at low thinking for routine steps, and keep an eye on the xhigh setting for the hard ones.

FAQ

What is Google Gemini 4 Hydrogen?

Google DeepMind’s fast execution model in the Gemini 4 family, launched November 12, 2026. Google calls it its daily driver for agentic workflows: everyday coding, multi-step agent tasks and multimodal synthesis, at $1.50 per million input tokens and $7.50 per million output tokens, with a 1 million-token context window.

How much does Gemini 4 Hydrogen cost?

$1.50 per million input tokens and $7.50 per million output tokens on the Gemini API, flat across the full 1 million-token context window, with cached input at $0.075 per million. There is also a free API tier with limits. Gemini 3.8 Flash stays cheaper ($0.75/$3.75) only until December 31, 2026; from January 1, 2027 its prices match Hydrogen’s.

Is Gemini 4 Hydrogen faster than Gemini 3.8 Flash?

Google says it completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking setting, and it scores 57% on Terminal-Bench 4.1 versus 18.2% for Flash. But the independent Artificial Analysis harness measured a much smaller gap in time per task (13.5 minutes versus 14.8), even though Hydrogen’s raw output speed is far higher (190 tokens per second versus 112). The 2.5x figure is Google’s claim, not an independent measurement.

Where can I use Gemini 4 Hydrogen?

It is rolling out globally in the Gemini API and Google AI Studio, Gemini for business, the Gemini app, Google Search, and for paid Google AI Pro and Ultra subscribers in Google Antigravity. It is the first Gemini 4 model available to free users.

How is Gemini 4 Hydrogen different from Gemini 4 Argon and Carbon?

Hydrogen is the fast, everyday execution tier, adapted from Carbon. It scores 51 on the Artificial Analysis Intelligence Index at its default high setting (54 at xhigh), one point above Argon at 50 and below Carbon at 60, and it costs less than both: $1.50/$7.50 per million tokens versus $2/$10 for Argon and Carbon’s introductory rates. Argon and Carbon are built for deep, long-horizon reasoning; Hydrogen is built for getting routine agentic work done quickly and cheaply.