ModelsArchitectures & capability
Google Rolls Out Gemini 4 Hydrogen to Search for All Users Worldwide
Google has made Gemini 4 Hydrogen, a cheaper model derived from Gemini 4 Carbon, the engine of Search for all users worldwide. It promises better intent understanding and broader query fan-out, while the frontier Gemini 4 models remain restricted.
From the Google AI article of November 12, 2026 and cited sources: On Nov 12, 2026, Google made Hydrogen live in Search worldwide with no paid plan required. [1] [5] [6] [8]
SOURCES[1] Google Search gets a major upgrade with Gemini 4 Hydrogen… blog.google[5] tbreak: Gemini 4 Hydrogen reaches Google Search for all… tbreak.com[6] heise: Gemini 4 Hydrogen, efficient alternative to Argon… heise.de
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- On Nov 12, 2026, Google made Hydrogen live in Search worldwide with no paid plan required. [1] [5] [6] [8]
- Original source
- Google AI article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 12, 2026 article published
- Covered
- Nov 13 edition
- Evidence
- 11 sources across 9 sites
- 1Google Search gets a major upgrade with Gemini 4 Hydrogen…blog.google
- 2Axios: Google unveils long-awaited Gemini 4axios.com
- 3Artificial Analysis: Gemini 4 Carbon new number 1artificialanalysis.ai
- 4Introducing Gemini 4 Hydrogen (Google DeepMind)blog.google
- 5tbreak: Gemini 4 Hydrogen reaches Google Search for all…tbreak.com
- 6heise: Gemini 4 Hydrogen, efficient alternative to Argon…heise.de
QuietHow it’s measured
Hacker News · 6 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
On Nov 12, 2026, Google made Hydrogen live in Search worldwide with no paid plan required. Google says it reads intent better, runs more parallel sub-queries across subtopics (query fan-out), links to more sites, and returns more visual answers. Fan-out already powered AI Mode, so Hydrogen refines that approach rather than introducing it. [1] [5] [6] [8]
Through the API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, about 75% of the list price for Argon and Carbon. It handles up to 1M tokens in and out and has four thinking levels. Artificial Analysis measured about $0.29 per task at low and $3.16 at xhigh, so the chosen thinking level largely sets the bill. [3] [4] [6]
Artificial Analysis scored Hydrogen 54 on its Intelligence Index at xhigh and 51 at high, close to Claude Sonnet 5.5 and GPT-6.1 Sol at 55. It recorded an 8% hallucination rate and tied for first on a document-analysis test. Hydrogen trailed rivals on Terminal-Bench 4.0 (54% versus 56–64%) and was among the most verbose models at xhigh. [6] [7]
Hydrogen is live and free in Search, and its API price is about 75% of Carbon's. Google's Search-quality claims carry no metrics, and Artificial Analysis tested the API model, not Search answers.
Read the full assessment
Implication: publishers should track referral data, and developers should benchmark terminal-heavy work.
I'm writing up the dossier from Google's two launch posts, Artificial Analysis results, and trade-press coverage.
Executive brief
The model behind Google's "biggest model upgrade to Search yet" is not its top Gemini 4 model. It's Gemini 4 Hydrogen, a faster and cheaper model derived from Gemini 4 Carbon, and it is now free to all Search users worldwide. Google says it improves understanding of intent, sends out more parallel sub-queries ("query fan-out"), links to more sites and gives more visual answers (Google). Google released no Search-specific metrics. The benchmark data covers the model through its API, not Search quality, and no independent evaluation of Search answers was found.
What changed and event timeline
Argon debuts to a small group
Google first gave Gemini 4 Argon to selected cybersecurity partners. Developers and consumers were to get access later.
Carbon tops an independent index
Artificial Analysis ranked Gemini 4 Carbon first on its Intelligence Index v5 with 59. It said Carbon and Argon were still limited to selected users.
Hydrogen launches across Google products
DeepMind's Koray Kavukcuoglu launched Hydrogen in the API, AI Studio, the Gemini app, Search, Sheets and other products.
Search rollout
Robby Stein, Google's VP of Product for Search, announced that Hydrogen is live in Search globally for all users.
More detail
Independent scores reported
Coverage of Artificial Analysis testing shows clear strengths and weaknesses at each reasoning setting.
More detail
Capabilities and access
- Model: Gemini 4 Hydrogen. The posts give no versioned API model ID.
- Search: live globally for all users, with no paid plan needed (tbreak). Language gains apply to all languages AI Mode supports (Google).
- API: $1.50 per million input tokens and $7.50 per million output tokens, the same rate across the whole context.
Read the full section
- Model: Gemini 4 Hydrogen. The posts give no versioned API model ID.
- Search: live globally for all users, with no paid plan needed (tbreak). Language gains apply to all languages AI Mode supports (Google).
- API: $1.50 per million input tokens and $7.50 per million output tokens, the same rate across the whole context. Up to 1M tokens in and 1M out. Four thinking levels: low, medium, high (default) and xhigh (Google DeepMind).
- Gemini app: free users get Hydrogen through automatic model selection (tbreak).
Technical analysis for researchers and developers
- What Google says changed in Search: better instruction following and tool calling, plus stronger reasoning that runs parallel sub-queries across subtopics (query fan-out) (Google).
- Lineage: Hydrogen is derived from Carbon (heise). No architecture details are published.
- Reasoning level drives cost: Artificial Analysis measured about $0.29 per task at low and $3.16 at xhigh, and 1.3 to 13.7 minutes per run (heise).
Read the full section
- What Google says changed in Search: better instruction following and tool calling, plus stronger reasoning that runs parallel sub-queries across subtopics (query fan-out) (Google). Query fan-out means splitting one question into several searches run behind the scenes (ContentGrip).
- Lineage: Hydrogen is derived from Carbon (heise). No architecture details are published.
- Reasoning level drives cost: Artificial Analysis measured about $0.29 per task at low and $3.16 at xhigh, and 1.3 to 13.7 minutes per run (heise). Which level Search uses is not disclosed.
- Model card: a model card exists. Its details are not covered here.
Claims and evidence
- Google's own claims: 2.5× faster than Gemini 3.8 Flash; 57% vs 18.2% on Terminal-Bench 4.1. Up to 81% less overreach in agentic tasks (Google DeepMind).
- Artificial Analysis results: Intelligence Index of 54 at xhigh, 51 at high, 41 at medium and 27 at low.
- Search quality: Google's Search claims ("more precise," "broader range of links") come with no numbers (Google).
Read the full section
- Google's own claims: 2.5× faster than Gemini 3.8 Flash; 57% vs 18.2% on Terminal-Bench 4.1; up to 81% less overreach in agentic tasks (Google DeepMind).
- Artificial Analysis results: Intelligence Index of 54 at xhigh, 51 at high, 41 at medium and 27 at low. An 8% hallucination rate on its knowledge tests, compared with 43% for Claude Sonnet 5.5. Tied for first on GDP.pdf, a document-analysis test (Trending Topics).
- Search quality: Google's Search claims ("more precise," "broader range of links") come with no numbers (Google).
Context and prior work
- Gemini 4 family: Argon (Sep 30) and Carbon (Oct 28) are the frontier models, both limited to selected users at first.
- What's new: query fan-out already powered AI Mode. Hydrogen is meant to make it work better, not to introduce it (ContentGrip).
- Rollout pattern: Google put the cheaper model in front of everyone while the frontier models stayed restricted.
Read the full section
- Gemini 4 family: Argon (Sep 30) and Carbon (Oct 28) are the frontier models, both limited to selected users at first. Both are listed at $2 per million input tokens and $10 per million output (Axios, Artificial Analysis, tbreak).
- What's new: query fan-out already powered AI Mode. Hydrogen is meant to make it work better, not to introduce it (ContentGrip).
- Rollout pattern: Google put the cheaper model in front of everyone while the frontier models stayed restricted.
Limitations, safety and contested findings
- Command-line weakness: 54% on Terminal-Bench 4.0, against 56–64% for competitors (heise).
- Verbosity: at xhigh it ranks second-to-last for verbosity among 77 models (Trending Topics).
- Conflicting scores: reports disagree on reference scores. Index versions may differ.
Read the full section
- Command-line weakness: 54% on Terminal-Bench 4.0, against 56–64% for competitors (heise).
- Verbosity: at xhigh it ranks second-to-last for verbosity among 77 models (Trending Topics).
- Conflicting scores: reports disagree on reference scores. Argon is put at 50 (Artificial Analysis) or 53 (Trending Topics search listing), and Carbon at 59 or 60 (Trending Topics). Index versions may differ.
- Publisher traffic: publishers worry that AI Mode cuts their click-throughs (Digiday). Google says Hydrogen surfaces more links, but it gives no traffic data.
Business and practitioner implications
- SEO and publishers: more sub-queries per question could put more niche pages into answers. More tables and custom tools could also mean fewer clicks.
- Developers: about 75% of Carbon's list price, and a 51–54 index score close to Claude Sonnet 5.5 and GPT-6.1 Sol at 55 (heise).
- Cost control: the thinking level changes cost per task by about 11×, so set it deliberately.
Read the full section
- SEO and publishers: more sub-queries per question could put more niche pages into answers. More tables and custom tools could also mean fewer clicks. Watch referral data rather than relying on Google's description.
- Developers: about 75% of Carbon's list price, and a 51–54 index score close to Claude Sonnet 5.5 and GPT-6.1 Sol at 55 (heise). That makes it a strong default for agents that read documents. Test terminal-heavy coding before switching.
- Cost control: the thinking level changes cost per task by about 11×, so set it deliberately.