AI News

Google Releases Gemini 4 Hydrogen and Integrates It into Search

cnBetaUpdated Nov 14 15:43

AI summary

Google has released Gemini 4 Hydrogen, which is now available to users worldwide and covers all languages supported by AI Mode. Google says that at the default High reasoning level, it completes tasks 2.5 times faster than the previous-generation Gemini 3.8 Flash, with more than 80% fewer fabricated facts; on the Artificial Analysis Intelligence Index, it scores 51 at the High level and 54 at the Xhigh level, surpassing Gemini 4 Argon's 50. API pricing is $1.5 per million input tokens and $7.5 per million output tokens, lower than Argon's $2 and $10; Google has also launched Orchestrator for Antigravity, which can automatically switch between Carbon and Hydrogen within the same task, initially available to Google AI Ultra users.

Why read this

SelectedAI score 80

The original text provides benchmark and pricing comparisons between Hydrogen, Argon, and Flash, allowing readers to assess the positioning shift of this generation of efficient models.

Topics

Tags

Body · AI translation

Google Search has also integrated Hydrogen, enabling it to better understand complex, fragmented, and even misspelled questions, and to provide more intuitive answers and more relevant webpage links.

According to Google, at the default High reasoning level, Hydrogen completes tasks 2.5 times faster than the previous-generation Gemini 3.8 Flash, with further improved instruction-following capabilities and more than an 80% reduction in fabricated facts.

On the Artificial Analysis Intelligence Index, it even surpasses the higher-positioned Gemini 4 Argon.

Robby Stein, Vice President of Google Search Products, called this integration the largest model upgrade to Google Search to date. Gemini 4 Hydrogen has begun rolling out to users worldwide, covering all languages supported by AI Mode.

Now, you can pour all the thoughts in your head directly into the search box. Even if the question is long, has many conditions, or even contains spelling errors, Hydrogen can more accurately understand what you are actually looking for and organize its answer according to those requirements.

Google says the upgraded search can choose a more appropriate presentation format based on the question. When information needs to be compared, it can generate tables; for complex knowledge questions, it can use images, visual content, or even interactive tools to help users understand the answer.

Things that previously required opening multiple websites and repeatedly cross-checking information can now potentially be resolved directly on the search page.

At the same time, Google Search has also upgraded its Query fan-out technology. When facing complex questions, the system can simultaneously search multiple related directions, cover a broader range of information sources, and provide more relevant webpage links.

This is actually quite interesting. AI search has long been criticized for leaving answers directly on the search page, making users increasingly less likely to click through to external websites. Now, however, Google is starting to emphasize that the upgraded AI search can help users discover more websites.

Google itself is also facing monetization pressure from AI search. On the October 28 earnings call, Google Chief Business Officer Philipp Schindler acknowledged that AI Mode's monetization efficiency is lower than traditional search, and that its rapid query growth has also dragged down the growth rate of search advertising revenue.

As for whether websites can actually regain more clicks after links increase, time will be needed to prove it.

Gemini is more obedient and more capable

Google may feel that the codename Hydrogen is intuitive enough. Although the name does not include Flash, Hydrogen plays the role of the high-efficiency model in the Gemini 4 series.

Google positions Carbon and Argon for complex deep reasoning tasks, while Hydrogen places more emphasis on execution efficiency, especially coding, tool calling, and multi-step tasks.

But this model, which focuses on speed and execution efficiency, is even more capable than its own high-end models.

On the Intelligence Index leaderboard of third-party evaluator Artificial Analysis, Hydrogen scores 51 at the High reasoning level and further rises to 54 at the Xhigh level, surpassing Gemini 4 Argon's 50.

According to Google's official benchmarks, Hydrogen performs especially prominently in knowledge work, automation, and reasoning.

In Agents' Last Exam v1.1, which examines AI's ability to handle complex knowledge tasks, it achieves 71.6%, surpassing Claude Sonnet 5.5's 60.4% and GPT-6.1 Terra's 69.8%.

In the Vending-Bench 2 test, which simulates long-term business operations, Hydrogen ultimately accumulated a fund balance of $14,549, also leading the other two models.

In coding, it leads with 13.5% in SWE-sweep, but in Terminal-Bench 4.1 it trails Claude Sonnet 5.5's 63.9% with 57%.

Beyond benchmark scores, the more direct change is that Gemini is more obedient and more capable in everyday use.

Google says Hydrogen can better follow tone, format, and specific rules, maintain task goals across multi-turn conversations, and reduce the number of repeated revisions.

When handling complex materials, it can extract information from uploaded files, organize tables and visual content, and complete tasks in conjunction with Gmail, Google Drive, and connected applications such as Airtable and Dropbox.

Hydrogen's tool-calling capabilities have also improved. Google says the model can more accurately select the tools it needs to use, and even if a tool call fails or a search returns no results, it can adjust subsequent actions.

In an official demo, Hydrogen used Google Antigravity to generate a multi-page slide deck directly in Google Slides based on a YouTube talk video and a design template. It then autonomously reviewed the presentation and fixed layout issues such as overlapping text.

There is another notable change: this time, Google launched a new Orchestrator feature for Google Antigravity.

Antigravity previously already supported multiple Agents working together using different models. The newly launched Orchestrator can, within the same task, automatically switch between Carbon and Hydrogen based on execution progress.

According to the workflow Google provided, Carbon can first analyze the task and formulate a plan, then hand it off to Hydrogen to modify code, update dependencies, and run tests. If new difficulties arise during execution, the task can be handed back to Carbon for analysis and adjustment.

The entire process does not require the user to manually switch models.

Google says that in internal testing, Orchestrator achieved a task success rate close to using Carbon alone, while reducing quota consumption by an average of 24% and increasing task completion speed by 19%.

This feature is currently being rolled out first to Google AI Ultra users, and will support more models and subscriptions in the future.

On pricing, Hydrogen's API charges $1.5 per million input tokens and $7.5 per million output tokens, lower than Gemini 4 Argon's current discounted prices of $2 per million input and $10 per million output.

Compared with the previous-generation Flash, the new model's price has increased. Gemini 3.8 Flash currently charges $0.75 for input and $3.75 for output, only half of Hydrogen's.

However, the discounted price of the old Flash will end on December 31, and starting January 1 next year it will be adjusted to the same level as Hydrogen.

Netizens: So what happens to Argon?

Right after Hydrogen was released, people immediately started testing it.

Netizen @SPAC89 showed a Three.js project they made with Hydrogen Extra. They said the entire project took less than two hours, of which nearly 80% of the time was spent on rendering, and it consumed only 5% of the weekly quota of a Google AI Pro subscription.

They were quite satisfied with the new model, and even said this was the first time they felt that the $20 Pro subscription had almost no quota limit in practice.

User @StatsWire ran a same-prompt test with four Gemini models: having them each use Three.js to create a playable 3D Mario game, without calling any external 3D models or textures.

In the comments, some netizens asked the author to also publish the time and cost required for each model to generate the game. The author said these data would be added in the next test.

Some netizens thought of a rather interesting scenario: "Developers in 2026: AI wrote a bug, now you throw the error message into Google Search, and the same AI is still the one explaining the bug."

These complaints are not aimed at Hydrogen, but are more about dissatisfaction with Google's existing products.

Under Google DeepMind's official launch post, Hydrogen's performance also raised another question: so what use is Argon now?

After all, Hydrogen is cheaper, faster, and in some evaluations has even surpassed the higher-positioned Argon.

Some also questioned the speed improvement—what is the point of just being faster?

One netizen pointed out that a 2.5x speed increase is indeed good, but if the answer is wrong, the cost of re-running the task may be higher than a slow but correct answer.

Netizen @mattyapro's comment expressed what many people were thinking:

"For me, a model I use every day is more important than the ones topping the leaderboards."

I'm glad everyday work is getting attention.

Source

cnBeta