Gemini 4 Hydrogen vs 3.8 Flash: faster, sharper, double the price

Hydrogen is the better model in almost every way Google has measured, and it costs more to run through the API. It is the new everyday, fast tier of the Gemini 4 family, sitting below the heavier Gemini 4 Carbon and Gemini 4 Argon models that Google aims at long, hard reasoning jobs.

• Speed: 2.5 times faster than 3.8 Flash at its default high setting, Google says.
• Coding: 57% on Terminal-Bench 4.1, up from 18.2% for 3.8 Flash. That test checks how well a model gets real work done in a command-line window, the text-only workspace developers use to run code.
• Accuracy: more than 5x fewer fabricated claims than 3.8 Flash.
• API price: $1.50 input and $7.50 output per million tokens, against $0.75 and $3.75 for 3.8 Flash.
• Context: a 1 million-token window, priced at one flat rate across the whole window.

A token is a short fragment of text, often part of a word, and models are billed by the token. You pay for both what you send in and what comes back, so these rates matter most to apps that send a lot of traffic, not to someone typing a few questions a day.

On one widely watched scoreboard, yes, and that is the strangest part of this launch. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 using its standard (high) reasoning setting, one point ahead of Argon’s 50. Push Hydrogen to its new xhigh setting and it reaches 54.

That doesn’t make Argon obsolete. Google still pitches Carbon and Argon for deep architectural decisions and subtle bugs, and Hydrogen for the quick, disciplined work of writing code, running tests and checking changes. Koray Kavukcuoglu, a senior vice president at Google DeepMind, described Hydrogen as a model that can work alone or act as the fast engine alongside the bigger two.

Google is also adding a tool called Orchestrator to Google Antigravity, its AI coding workspace, which swaps between Carbon and Hydrogen inside a single session. On Terminal-Bench 4.0, Google says that mix came close to Carbon’s success rate while using 24% less quota and finishing 19% faster. In plain terms: let the expensive model plan, and let the cheap, quick one do the legwork.

The four thinking levels

Hydrogen lets you choose how hard it thinks, from low up to the new xhigh, with high as the default. Lower levels answer faster and cost less; higher ones spend more time reasoning. One detail matters for anyone migrating code: the minimal thinking level and the older thinking_budget setting are not supported on this model, so apps tuned for 3.8 Flash may need a small rewrite.

For most people, Hydrogen arrives without lifting a finger. It now powers the AI features in Google Search for all users worldwide, a job 3.8 Flash had done since it launched in September. Google says Search answers will look more visual, show a wider range of links and cope better with chatty, typo-filled questions.

In the Gemini app, Hydrogen shows up in the model picker. Free users don’t always get to pick it by hand; the app hands a prompt to Hydrogen when it decides the model is the right fit. Google says the new model sticks more closely to your tone and format, so email drafts, study guides and rewrites should need fewer second attempts. It is also meant to be better at choosing the right tool, such as pulling details from your Gmail or Google Drive when you ask about them.

• Free Gemini app users: Hydrogen is used automatically for some prompts.
• Google AI Pro and Ultra subscribers: access in the app and in Antigravity, with Gemini in Google Sheets getting Hydrogen first on these plans.
• Businesses: available through the API or the model menu in Gemini Enterprise.
• Developers: the Gemini API, Google AI Studio across all tiers, Android Studio and Vertex AI.

That is a big change from how Google handled Carbon and Argon. Those rollouts were staged, started with cybersecurity researchers and, even at general release, were limited to people paying for Google’s most expensive AI plan.

For many apps, probably, but run your own numbers before you switch. Doubling the per-token rate hurts most when your app produces long answers, because output tokens cost five times as much as input tokens. Thinking tokens, the hidden reasoning a model does before answering, are billed as output too, so cranking Hydrogen up to xhigh will cost more on every request.

Some of that gets clawed back. In early partner tests on graph, SQL and data-heavy enterprise work, Google says Hydrogen ran 2x to 2.8x faster than 3.8 Flash and used 12% to 23% fewer tokens. Fewer retries also matter: a model that invents fewer facts and recovers when a tool call fails can save whole rounds of back-and-forth. Those savings won’t cancel out a doubled rate on their own, though. A good test is to replay a day of real traffic at the low and medium levels and compare the bill with what you pay now.

There are cheaper routes. Google offers Batch and Flex processing at different rates for work that doesn’t need an instant answer, plus a limited free tier in AI Studio and the API. Read the free tier’s terms first: Google says it can use what you send on the free tier to make its products better, a use it rules out for paying customers. Hydrogen supports automatic (implicit) caching of repeated context, but not explicit caching yet, which matters if your app resends the same long documents.

Google says Hydrogen is up to 2x less likely than 3.8 Flash to work around a permission boundary when a task is blocked, and that it cuts what the company calls overreach in agent tasks by up to 81%. Those are useful gains for anyone letting an AI agent touch files or accounts.

Google is also clear that the model still gets things wrong. Like every chatbot, it can hallucinate, so check anything important it tells you, especially numbers, quotes and medical or legal details. The company says work on jailbreak resistance is continuing.