What Is Gemini 4 Hydrogen? Google’s Fast Model for Coding and Agents, Explained
Gemini 4 Hydrogen is Google’s fast, affordable model for everyday coding and agentic work — the third model in the Gemini 4 family, launched on November 12, 2026, after Argon and Carbon. Where Carbon is Google’s strongest model for the hardest, longest engineering jobs, Hydrogen is the one built to run all day: writing code, running tests, verifying changes, carrying multi-step tasks through to completion, and doing it quickly and cheaply. It costs $1.50 per million input tokens and $7.50 per million output tokens, it has a free tier in the Gemini API, and it is rolling out to everyone — including every Google Search user.
This guide explains what Hydrogen is, what it is best at, what it costs, how to access it, and how it compares with Carbon, Argon and the model it replaces, Gemini 3.8 Flash. All benchmark figures below are vendor-reported unless stated otherwise: they are Google’s own numbers, on benchmarks Google chose, and no outside lab had independently reproduced them at launch.
What is Gemini 4 Hydrogen?
Hydrogen is a large language model from Google DeepMind, and the third member of the Gemini 4 series. Argon arrived on October 13 as the deep-reasoning flagship, Carbon followed on October 28 as the strongest model in the range, and Hydrogen completes the family from the other direction: not the cleverest model, but the fastest and cheapest one that still carries what Google calls cutting-edge intelligence.
Google’s pitch is that most real work does not need a frontier model. Teams still need Carbon or Argon for deep architectural judgement and untangling subtle bugs, but the bulk of everyday development — writing code, running tests, verifying changes — needs a fast, disciplined execution model instead. Hydrogen is built for that role: Google describes it as its daily driver for agentic workflows, the model you run constantly while the bigger models are called in only for the hardest steps.
That idea shapes how Google is selling it. In Google Antigravity, its agent-first development platform, a new feature called Orchestrator switches dynamically between models inside a single session — Carbon for architecture, coordination or audit work, Hydrogen for the implementation steps — and Google reports that this pairing reaches near-Carbon success rates on Terminal-Bench 4.0 while consuming 24 percent less quota and finishing 19 percent faster. For background on how these systems work, see our guide to agentic AI.
Hydrogen has a 1 million-token input context window, and Google says it can produce up to 1 million output tokens — the same input window as its siblings, and enough to hold large repositories and long documents in full before work starts.
What Hydrogen is best at
Everyday coding and debugging
Hydrogen’s home ground is iterative repository work: resolving issues, debugging, rapid iteration and answering questions about a codebase, with focused, surgical edits that fix the problem without rewriting unrelated files. Google says it outperforms similar-tier models on SWE-sweep, a test where the model has to discover and repair bugs without breaking existing code, and that it is good at batching file-read and search calls in parallel rather than inspecting files one at a time.
The upgrade over the model it replaces is large. Against Gemini 3.8 Flash, Google reports that Hydrogen finishes tasks 2.5 times faster at its default high thinking setting, more than triples its Terminal-Bench 4.1 score from 18.2 percent to 57 percent, and cuts fabricated or unsupported claims by more than five times. In early testing with partners on complex graphs, SQL and data-heavy enterprise workflows, Google says Hydrogen ran two to 2.8 times faster while using 12 to 23 percent fewer tokens than 3.8 Flash.
Open-ended agentic execution
On multi-step work such as deep research, data analysis and broad codebase navigation, Hydrogen ranks fifth on Agent Arena, the live benchmark run by arena.ai. For deep domain reasoning it leads its class on HLE-Diamond at 58.8 percent, ranks first overall on the Vals Tax Agent Bench at 52.3 percent, ties for first on the Vals Legal Research Bench at 55.8 percent, and leads its class on AutomationBench-AA, which tests whether agents can complete assignments across real business applications.
These are narrow, specialist tests, and a first place on a tax-agent benchmark does not make Hydrogen the best model for every job. What the pattern does show is where Google aimed it: at professional, multi-step work where the model has to use tools, check sources and produce a finished deliverable, not at single-turn chat.
Multimodal knowledge work
Hydrogen also turns dense PDFs, spreadsheets, charts, audio and recorded video into cited reports and structured deliverables. Google reports that it outperforms other models on the Multimodal Knowledge Agent benchmark at 53.4 percent and scores 71.6 percent on Agents’ Last Exam v1.1, which tests end-to-end professional workflows — the model for turning a pile of messy source material into a finished document.
Web development and everyday help
On live UI prototyping, design-to-code iteration and interactive app building, Google says Hydrogen produces more functional, feature-complete layouts in fewer prompts than its predecessor — the kind of speed that matters when a developer is iterating in front of a client rather than running an overnight job. That is a different kind of speed from Gemini Omni 1.2 Flash, launched the same week as Carbon, which is a fast multimodal model built for short video and everyday tasks.
For consumers, Hydrogen also upgrades the Gemini app and Google Search. In the app, Google says it follows instructions more closely across multi-turn conversations, calls the right tools more often, and pulls context from uploads, Gmail, Google Drive and connected apps such as Airtable and Dropbox. In Search, it is rolling out to all users globally, with better understanding of multi-part and mistyped queries, more visual responses, and a wider range of relevant links. Google’s Search team calls it the biggest model upgrade to Search yet.
Four thinking levels: low, medium, high, xhigh
Hydrogen ships with four configurable thinking levels — low, medium, high (the default) and a new xhigh tier — so developers can choose how much reasoning a task gets, and pay only for the depth it needs. A quick code completion can run on low; a tangled debugging job can go to xhigh.
The levels scale in a smooth, predictable progression, according to Google. At its default high setting, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, one point ahead of Argon at 50; at xhigh it reaches 54. That is a striking result for a model priced below both of its siblings, though it needs the usual caution: a one-point lead over Argon does not mean Hydrogen is the stronger model overall. Carbon, for comparison, scored 60 on the same index when it was checked on November 10.
One practical warning belongs here. Google’s own model card notes that higher thinking levels can use more tokens, and can occasionally be slower or time out.
How much does Gemini 4 Hydrogen cost?
Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1 million-token context window. There is no price step for longer contexts, which is a genuine simplification: many rival models charge more once a prompt passes a threshold, and Hydrogen does not.
Unlike Argon and Carbon, Hydrogen also has a free tier in the Gemini API, listed on Google’s Gemini API pricing page. That makes it the first Gemini 4 model that developers can try without paying, and the first one free-tier users can reach at all.
The price needs context. At $1.50 and $7.50, Hydrogen costs twice what Gemini 3.8 Flash costs today — but it is exactly what 3.8 Flash will cost from January 1, 2027, when Google raises that model’s price. In other words, Hydrogen is not a cheaper Flash; it is Flash’s replacement at Flash’s future price, with substantially better performance. Against its own family, Hydrogen undercuts both siblings: Argon and Carbon both launched at $2 per million input and $10 per million output tokens. Against Anthropic’s Claude Fable 5.2 at $8 and $40, Hydrogen costs roughly a fifth — though Fable 5.2 still leads on the broad Artificial Analysis index, so the comparison is price against capability, not price alone.
How to access Gemini 4 Hydrogen
Hydrogen rolled out globally on November 12, 2026, and it is the most widely available Gemini 4 model yet:
- Developers: the Gemini API and Google AI Studio on all tiers — including the free tier — plus the paid tier in Android Studio. The developer companion post points to the API guide for setup and model details.
- Businesses: Gemini for business, through the API or the model picker, plus Google Antigravity and Gemini in Google Sheets for eligible business subscriptions.
- Consumers: the Gemini app, through the model picker, and Google Search for all users — the first time a Gemini 4 model has reached free Search users. In Antigravity, Hydrogen is available to paid Google AI Pro and Ultra subscribers.
One nuance for free users of the Gemini app: Hydrogen reaches them through automatic model selection rather than a direct picker choice. Google has not published a country-by-country rollout schedule beyond the global launch, and availability in specific products can still vary by region.
Hydrogen vs Carbon vs Argon vs 3.8 Flash
The Gemini 4 family now has three models, and the choice between them is about the job, not a single leaderboard. For the full picture on the strongest model, see our guide to Gemini 4 Carbon.
| Model | Price ($/1M input / output) | Artificial Analysis index | Best for |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 / $7.50, free tier | 51 (high), 54 (xhigh) | Everyday coding, agentic execution, Search |
| Gemini 4 Argon | $2 / $10 | 50 | Deep reasoning, most engineering work |
| Gemini 4 Carbon | $2 / $10 (intro, then $4 / $20) | 60 | Long-horizon engineering, cybersecurity |
| Gemini 3.8 Flash | $0.75 / $3.75 (rising to $1.50 / $7.50 in Jan 2027) | — | Being replaced by Hydrogen |
Hydrogen is the cheapest and fastest of the three, and the only one with a free tier — but it is not the strongest. Carbon leads the family on the broad index by nine points over Hydrogen at its default setting, and on the hardest agentic tests: 65.2 percent on Terminal-Bench 4.0 against Hydrogen’s 57 percent on the newer 4.1 test, and state-of-the-art results on every cyber benchmark Google cites. Argon sits between them on capability, and Google still recommends it for most everyday engineering work where depth matters more than speed.
Against rivals outside Google, the picture is similar. Claude Fable 5.2 leads the Artificial Analysis index at 65 points at maximum effort, well ahead of Hydrogen’s 54 at xhigh, and leads on coding benchmarks such as FrontierCode 2.0 — but it costs more than five times Hydrogen’s token price. And as with every model, the token price is not the job price: a model that finishes in one attempt can be cheaper than one that needs three retries, so measure cost per finished job, including retries and review time.
The practical rule is simple. If the work is long, security-sensitive or architecturally hard, use Carbon. If it needs deep reasoning but not Carbon’s price or power, use Argon. If it is everyday coding, research, document work or agentic execution where speed and cost matter, Hydrogen is now Google’s default answer — and the one most people will actually meet, because it is the model inside Search.
What Google has not said
A careful reader of the launch materials will notice the gaps. Google has published no standalone latency figures beyond the claim that Hydrogen is 2.5 times faster than 3.8 Flash, and no country-by-country rollout schedule beyond the global launch. It has not said when Hydrogen will reach Vertex AI, the cloud platform many enterprises buy through, or whether the free API tier’s rate limits will suit production workloads.
The model card is more candid than the launch post. It lists hallucinations, occasional slowness or timeouts, and potentially greater token use — especially at higher thinking levels — among the model’s limitations, and it gives a March 2026 knowledge cutoff, so Hydrogen’s knowledge of later events comes only from tools and retrieval, not from training.
None of this is unusual for a launch, but it is the reason the benchmark figures in this guide are labelled vendor-reported throughout: the independent evaluations that will test Google’s claims have not had time to publish yet.
Who should use Gemini 4 Hydrogen?
Hydrogen makes most sense for four groups. Developers who want a fast, cheap model for everyday coding — issue resolution, debugging, rapid iteration — and who are migrating from 3.8 Flash, for whom it is the direct upgrade path. Teams building agentic workflows who need an execution engine that can run all day without the token bill of a frontier model, especially paired with Carbon as the orchestrator in Antigravity. Knowledge workers who need to turn dense PDFs, spreadsheets and videos into cited reports. And anyone who uses Google Search, where Hydrogen is now the model behind the answers, free or not. It makes less sense for the hardest jobs — multi-day engineering, security-sensitive work and deep architectural reasoning — where Carbon or Argon are the right models.
Frequently asked questions
Is Gemini 4 Hydrogen free?
Partly. The Gemini API has a free tier for Hydrogen, the first free tier for any Gemini 4 model, and it is rolling out to all Google Search users at no cost. The paid API price is $1.50 per million input tokens and $7.50 per million output tokens.
When was Gemini 4 Hydrogen released?
Google announced Hydrogen on November 12, 2026, with global availability the same day across the Gemini API, Google AI Studio, the Gemini app, Google Search, Gemini for business and — for paid Pro and Ultra subscribers — Google Antigravity.
How is Hydrogen different from Carbon and Argon?
Hydrogen is the third and fastest model in the Gemini 4 family, built for everyday coding and agentic execution rather than deep reasoning. Carbon is the strongest model, for long-horizon engineering and cybersecurity; Argon sits between them on capability. Hydrogen is cheaper than both — $1.50 and $7.50 against $2 and $10 — and the only one with a free tier, but it scores lower on the broad Artificial Analysis index: 51 at its default setting against 50 for Argon and 60 for Carbon.
Is Hydrogen better than Gemini 3.8 Flash?
According to Google, yes, by a wide margin: 2.5 times faster at its default setting, 57 percent against 18.2 percent on Terminal-Bench 4.1, and more than five times fewer fabricated claims. It costs twice Flash’s current price, but exactly what Flash will cost from January 1, 2027 — Hydrogen is Flash’s replacement, not a cheaper alternative.
Can I run Hydrogen on my own servers?
No. Hydrogen is proprietary and hosted only, distributed through Google’s API and products. There are no downloadable weights.
