Connect with us

NEWS

Google Puts Gemini 4 Hydrogen in Search for Everyone

Gemini 4 Hydrogen is live in Search for every user at $1.50 per million input tokens, outscoring Argon while Carbon stays the paid planner.

Published

on

Google launched Gemini 4 Hydrogen on November 12, a $1.50-per-million-token model now live in Search and the Gemini app for every user. Koray Kavukcuoglu, SVP of Google DeepMind and chief AI architect, said the company is bringing Gemini 4 to everyone with a model built for rapid execution, not long-horizon reasoning.

The sticker is the easy headline. The harder move is that Hydrogen is the Gemini 4 model Google can actually ship into Search, then hand the leftover hard cases to Gemini 4 Carbon through a new Orchestrator switch in Antigravity.

Hydrogen Is the Gemini 4 Model Everyone Can Open

Paid API access is priced flat across a 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens. That rate does not rise with prompt length. Kavukcuoglu set Hydrogen against Gemini 4 Carbon and Gemini 4 Argon, the two frontier models Google still reserves for deep, long-horizon work.

Sundar Pichai, CEO of Google and Alphabet, said Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash and carries four thinking levels so users only wait, and pay, for the depth a job needs. He named the same surfaces Kavukcuoglu did: all tiers of Google AI Studio, the Gemini API, the Gemini app, Search, and paid Google AI Pro and Ultra plans in Google Antigravity.

The Gemini app account later added a practical detail for people who never touch an API key. Free users get Hydrogen through automatic model selection when the app judges it the right model for the prompt, rather than through a picker on every plan.

HYDROGEN AT A GLANCE

  • API price: $1.50 per million input tokens and $7.50 per million output tokens, billed flat across the 1 million-token window.
  • Thinking levels: low, medium, high as the default, and a new xhigh tier.
  • Speed claim: 2.5 times faster than Gemini 3.8 Flash at the high default setting.
  • Consumer access: Search and the Gemini app for all users; Antigravity for paid Pro and Ultra, with a later expansion promised.

Varun Mohan, who is building Antigravity at Google DeepMind, said Hydrogen is launching to all paid Antigravity users and will expand to more users over time. That is the split that matters. The model that Google is willing to put in front of a Search box is not the same product as the one sitting in a coding IDE.

A One-Point Lead Over Argon at Default Thinking

Google’s own ranking claim is the one that will follow Hydrogen around. At its default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, against 50 for Argon. At xhigh it reaches 54. Carbon remains the stronger Gemini 4 model; Google did not publish Carbon’s index score in the Hydrogen post.

Artificial Analysis, the independent tester behind that index, put Hydrogen’s xhigh result one point below Claude Sonnet 5.5 and GPT-6.1 Sol, both at 55 on max effort. The same note said high reasoning matches Argon’s intelligence at about 19% of Argon’s cost per task, and that xhigh comes in at about 32% of Sonnet’s cost per task, while still costing about 2.5 times as much as Sol.

INDEPENDENT INDEX AND TERMINAL SCORES

Model Setting Intelligence Index Terminal-Bench 4.0
Gemini 4 Hydrogen high (default) 51 54%
Gemini 4 Hydrogen xhigh 54 54%
Gemini 4 Argon high 50 Not listed
Claude Sonnet 5.5 max 55 64%
GPT-6.1 Sol max 55 56%

Artificial Analysis also scored Hydrogen at 8% on its AA-Omniscience hallucination measure, against 43% for Sonnet 5.5 at max and 52% for Sol at max. Google’s own comparison with 3.8 Flash is older and narrower: Hydrogen cuts fabricated claims by more than 5 times and more than triples Terminal-Bench 4.1, at 57% versus 18.2%.

Those two terminal figures are not the same test. Google’s 57% is Terminal-Bench 4.1. Artificial Analysis published 54% on Terminal-Bench 4.0, and said xhigh does not move that 4.0 score at all. Hydrogen leads Google’s class comparisons on HLE-Diamond at 58.8%, ranks first on the Vals Tax Agent Bench at 52.3%, ties for first on the Vals Legal Research Bench at 55.8%, and scores 53.4% on Google’s Multimodal Knowledge Agent test. It sits fifth on Agent Arena’s live leaderboard.

Search Gets Hydrogen on Day One

Argon and Carbon did not land this way. Hydrogen is in Google Search globally for all users on the day of launch. Robby Stein, VP of product for Google Search, called it the biggest model upgrade to Search yet.

Stein’s post is the consumer version of the same architecture bet. Google is not asking anyone to pick Carbon before they can ask a messy question. It is putting the fast model under the box people already type into, then promising that the box will fan the question out across more of the web.

WHAT CHANGES IN SEARCH

  • Intent: Search is meant to follow multi-part questions, hard typos, and very specific requests in a natural train of thought, including across the languages AI Mode already supports.
  • Fan-out: Hydrogen runs simultaneous subtopic queries so results can point to a wider set of sites than a single rewritten query would find.
  • Layout: Answers are supposed to come back with more images, comparison tables, structured blocks, and custom tools, at Search speed.

That is a distribution choice, not a benchmark one. A model that is good enough to sit under every Search query does not need to win Terminal-Bench. It needs to be cheap enough, fast enough, and safe enough to run at Google’s volume, which is why the 51-versus-50 gap on Artificial Analysis is useful to Google even if it is only one point.

How Orchestrator Splits Planning From Typing

In Antigravity, Google shipped a model-selector option called Orchestrator that keeps Carbon or Argon in the planning seat and sends Hydrogen to touch the files. The Antigravity team said a static router that guesses difficulty from the first prompt breaks down once an agent runs long enough to need different models at different stages.

On Terminal-Bench 4.1, Orchestrator reached near-Carbon task success while consuming 24% less quota on average and completing 19% faster. Mohan repeated those two figures on X the same afternoon.

We’re also launching Orchestrator in the Antigravity dropdown which allows both Carbon and Hydrogen to work to complete your tasks. We’ve found that it consumes 24% less quota and gets the same tasks done 19% faster.

Varun Mohan, Antigravity, on X

HOW A TURN ACTUALLY MOVES

  • Planner first: Carbon or Argon opens in a read-only pass, maps dependencies, and writes a plan with triggers for when to re-plan.
  • Hydrogen executes: The cheaper model applies edits, updates packages, runs builds and tests, and checks the result against the plan.
  • Hand-back: Control returns to the planner if verification fails twice, an interface behaves unlike the plan, or a design choice was never settled.
  • Review: On longer jobs, Hydrogen reports progress so the planner can close the task, continue, or rewrite the plan.

Those hand-offs sit inside a single response. You do not switch models by hand. Orchestrator also runs a fixed configuration, so you cannot pick a thinking level for it. If you want xhigh on Hydrogen, or high on Carbon, you select that model directly.

Carbon Writes the Plan, Hydrogen Touches the Files

Google’s own example is blunt. Carbon starts by reading. Hydrogen writes the patch and runs the tests. If Hydrogen gets stuck, Carbon inspects state and repairs the plan. The expensive model is not the one burning tokens on package bumps.

That design has a cost that does not show up in the 24% figure. A session that is only reading a repository still opens on the planner, so Orchestrator does not make those turns faster. For that work, a Hydrogen-only pass, or a swarm of Hydrogen workers, is the shorter path. Orchestrator is a bet on mixed jobs, the ones that need a plan and then a lot of typing.

Orchestrator Stays on Paid Antigravity Plans

Access is narrower than Search. Orchestrator (Carbon) pairs Carbon with Hydrogen and is listed for Google AI Ultra. Orchestrator (Argon) pairs Argon with Hydrogen and is listed for Ultra and for paid Google AI Pro, including some device trials. Enterprises can reach both through Gemini for business. Hydrogen itself is on paid Pro and Ultra in Antigravity on day one, with Mohan saying the user base will widen later.

Google also told developers they can wire the same split themselves with the Antigravity SDK, putting Hydrogen on triage and UI iteration while Carbon takes refactors, new features, and security checks. The dropdown is the productized version of a pattern teams were already trying to build by hand.

December 1 Doubles the Price of Argon and Carbon

Hydrogen’s rate is the one Google did not put on a clock. Argon and Carbon still sell at an introductory $2 per million input tokens and $10 per million output tokens. After November 30, both move to $4 and $20. Hydrogen stays at $1.50 and $7.50.

From December 1, Hydrogen’s input price is 37.5% of the list rate on Argon and Carbon. The paid tier price on Google AI Studio also bills Hydrogen thinking tokens as output, which is where a high or xhigh pass can still get expensive even at the lower sticker.

GEMINI 4 LIST PRICES

Model Input per 1 million tokens Output per 1 million tokens From December 1
Gemini 4 Hydrogen $1.50 $7.50 Unchanged
Gemini 4 Argon $2.00 $10.00 $4.00 input, $20.00 output
Gemini 4 Carbon $2.00 $10.00 $4.00 input, $20.00 output

Argon arrived on September 30. Carbon went to a Fairwind cybersecurity group on October 28 and to paying developers on November 10, at the same introductory rate as Argon. Kavukcuoglu has been clear that Carbon uses more tokens than Argon to explore, reason, and verify. Hydrogen is the first Gemini 4 model with a free API tier, which is the other half of the same squeeze.

Once the December 1 step hits, a team that leaves Carbon on for everyday edits is paying for a model Google itself describes as the one you use when a task tests the limits. Orchestrator is how Google would like that team to keep Carbon’s plan quality without paying Carbon prices for every test run.

Sonnet 5.5 Still Owns the Terminal Jobs

Hydrogen’s weak spot is the work Google is selling it for. On Artificial Analysis, Sonnet 5.5 at max still leads Terminal-Bench 4.0, 64% to Hydrogen’s 54%. GPT-6.1 Sol sits at 56% on that test and at 58% on Terminal-Bench-Science 0.1, where Hydrogen reaches 50% only at xhigh, up from 41% at high. Google’s own Terminal-Bench 4.1 table, the one with Hydrogen at 57%, still has Sonnet 5.5 ahead at 63.9%.

Early coding use has the same grain. Hydrogen is already the daily driver for speed and quota, and it is the first Gemini 4 model that feels cheap enough to leave on. For messy branch work and versioned repos, Claude still looks more careful, and Hydrogen can sound finished when the patch is not. Google’s partner notes are more glowing: Box saw about twice the speed and 12% fewer tokens than 3.8 Flash, and Google said other early testers saw 2 times to 2.8 times faster runs with 12% to 23% fewer tokens on graph, SQL, and data-heavy jobs.

None of that undoes the Search decision. Free Gemini app users will meet Gemini 4 through automatic model selection, and Search users will meet it without opening a model menu at all. Carbon remains the model Google wants on the hardest planning loop. Hydrogen is the model it was willing to put under everyone else’s cursor.

Frequently Asked Questions

How Much Does the Gemini 4 Hydrogen API Cost?

The paid standard rate is $1.50 per million input tokens and $7.50 per million output tokens, and output includes thinking tokens. Google’s price list also posts a Batch and Flex pair at $0.75 and $3.75, a Priority pair at $2.70 and $13.50, and implicit cached input at $0.075. A free API tier exists, with rate limits, and Google says paid-tier content is not used to improve its products.

What Thinking Levels Does Gemini 4 Hydrogen Include?

Hydrogen ships four levels, low, medium, high, and xhigh, and defaults to high. In the Antigravity model menu the fourth level is labeled Extra, which maps to the xhigh budget; the CLI flags are –model gemini-4-hydrogen-xhigh or –effort xhigh. Orchestrator does not let you set a thinking level at all. Argon and Carbon still stop at low, medium, and high.

What Is Orchestrator in Google Antigravity?

Orchestrator is a selector option, not a separate model, and it is not a router that classifies your first prompt and sends the whole job to one SKU. Orchestrator (Carbon) is Ultra-only. Orchestrator (Argon) is on Ultra and paid Pro. If the user question is only a review or explanation, the planner can answer it without calling Hydrogen, and Hydrogen hands back after a verification step fails twice.

Is Gemini 4 Hydrogen Free in Search and the Gemini App?

Yes. Stein said Hydrogen is available globally in Search for all users, and the Gemini app says free users receive it through automatic model selection when the prompt fits. The free Gemini API tier is separate and rate-limited. Antigravity access on day one is still paid Pro, Ultra, and enterprise, with Mohan saying that pool will grow later.

Does Gemini 4 Hydrogen Support Context Caching?

Only implicit caching. Google’s model notes say Hydrogen does not support explicit cachedContents or caches.create calls. Implicit caching is on by default for prompts of at least 4,096 tokens and discounts cache hits by 95%, which is the $0.075 cached-input rate. Batch cached input is listed at $0.0375 per million tokens.

Harry is the editor and lead writer of STUDIO ONE NETWORKS, an independent title he owns and runs himself. Ten years in journalism, reporting first and editing later, taught him that entertainment and business are one beat seen from two sides: a box office figure is a company number, a streaming deal is a contract, a casting rumour is not a story until someone puts their name to it. He works from the record, whether that is a distributor's statement, a licensing agreement, an interview transcript or a set of published ratings, and checks every number against it before publication. The same rule holds for the rest of the site, which covers news, technology, science, sports, lifestyle, travel, auto and gaming for an audience spread across the world. When he gets something wrong, the article is corrected and the change is noted and dated, under a corrections policy anyone can read. Reader mail is answered by him at support@studioonenetworks.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending