Google Launches Gemini 4 Hydrogen Across Search and API | Zeniteq
Loading
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Google Launches Gemini 4 Hydrogen Across Search and API
Google’s November 12 model release brings faster agentic execution to its major products, with flat API pricing and a shared one-million-token context window.
Google released Gemini 4 Hydrogen on November 12, 2026, designed for rapid, agentic execution across Search, developer tools, and business products. Priced at $1.50 per million input tokens and $7.50 per million output tokens, it features a shared one-million-token context window. While Hydrogen excels at coding and parallel task execution, it is optimized for speed rather than deep reasoning. Developers should leverage its configurable thinking levels and integrate it as a worker model alongside more capable counterparts like Carbon, ensuring cost-efficient workflows by routing bounded, repetitive tasks to Hydrogen while reserving complex architectural decisions for higher-tier models.
Google introduced Gemini 4 Hydrogen on November 12, 2026, positioning it as a fast execution model for everyday coding, research and agentic workflows. The release spans Google Search, the Gemini app, developer tools and business products, giving Hydrogen a much broader role than an API-only model launch.
Google’s original announcement carries a November 16 update, but the launch date remains November 12. This is a distinct Gemini 4 model release, with its own pricing, model card and deployment across Google products.
Hydrogen’s place in the family is straightforward: Google positions Carbon and Argon for deep, long-horizon reasoning, while Hydrogen handles rapid execution. Developers can use it independently or assign it work inside a multi-model system.
Two technical details deserve particular attention. Its API pricing stays flat across the advertised context window, but that window is a combined input-output budget. And its model card identifies Hydrogen as an adaptation of Carbon, without explaining the adaptation technique.
Search Gives Hydrogen a Global Audience
Google says Hydrogen is available globally in Search for all users. The Search announcement focuses on four changes: better understanding of complicated requests, improved query fan-out, more visual responses and richer language understanding across supported AI Mode languages.
Query fan-out is the most consequential mechanism here. Search breaks a complicated question into related subtopics and runs simultaneous queries to gather relevant information. Google says Hydrogen improves how Search understands those subtopics and connects people to useful websites.
Consider a travel request combining transport, accessibility, budget and opening hours. Those constraints require several kinds of information. Better decomposition could help Search gather evidence for each part before assembling a response, although Google’s launch post does not provide independent measurements of answer quality or link coverage.
Google also promises more images, comparison tables, structured answers and custom tools. These changes affect how results are presented as well as which information Search retrieves.
Google Antigravity for paid Google AI Pro and Ultra subscribers.
The developer announcement also names Android Studio as an entry point. Availability across these products should not be read as identical access conditions: Google specifically attaches the Pro or Ultra subscription requirement to Antigravity access in its launch materials.
Flat API Pricing Makes Large Requests Easier to Budget
Google lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, with flat rates across its entire one-million-token context window.
API usage
Published price
Input
$1.50 per million tokens
Output
$7.50 per million tokens
Context-length pricing
Same rates across the advertised 1M-token window
These are Google’s published launch prices. They make the arithmetic predictable when a request contains a large repository or document collection: the stated per-token rate does not increase merely because the context gets longer.
For example, a request using 800,000 input tokens and 20,000 output tokens would cost:
Input: 0.8 × $1.50 = $1.20
Output: 0.02 × $7.50 = $0.15
Combined token charge: $1.35
That is an illustrative calculation from the published rates, not a measured workload. A coding agent may make many requests, send revised context repeatedly and invoke other services, so a complete task can cost substantially more than one exchange.
Output also deserves close attention: Hydrogen’s published output rate is five times its input rate. For applications producing long reports or extensive code, controlling unnecessary generation can matter even when input is inexpensive.
The Million-Token Window Is Shared
The model card describes a combined input-output context window of up to one million tokens. It lists text, images, audio and video as supported inputs.
The output limits have another important distinction:
Default synchronous output: up to 64,000 tokens.
Background long decode: up to one million output tokens.
Unified context: up to one million tokens shared between input and output.
Those figures should not be interpreted as one million input tokens plus another million output tokens in an ordinary request. The model card defines a shared budget, and its largest output allowance uses a separate background decoding path.
I’d plan integrations around that model-card wording. A repository that nearly fills the context leaves less room for generated material, while an application expecting a very long response needs to account for the difference between synchronous generation and background decoding.
Flat pricing simplifies the rate calculation. It does not remove context allocation or output-limit planning.
Four Thinking Levels Add a Second Cost Decision
Hydrogen supports four configurable thinking levels, according to Google’s developer announcement. The main launch post identifies high as the default and xhigh as the highest setting.
Google reports these Artificial Analysis Intelligence Index results:
Hydrogen at its default high setting: 51.
Hydrogen at xhigh: 54.
Gemini 4 Argon in the cited comparison: 50.
The performance and efficiency figures in Google’s launch materials are vendor-reported; the supplied publications do not provide independent replication of those results.
The practical choice is between spending more computation on a response and keeping execution fast. Google’s model card warns that higher thinking levels can use more tokens. It also lists occasional slowness, timeouts and hallucinations among the model’s limitations, and gives a March 2026 knowledge cutoff.
A higher thinking setting therefore needs evaluation against a specific task. Raising an aggregate intelligence score from 51 to 54 does not tell a team whether its bug-fixing agent will complete more accepted changes per dollar.
For routine edits with strong automated checks, I’d first evaluate the default setting. More demanding settings become useful when their additional reasoning produces enough improvement to justify the extra time and token consumption.
Google’s Benchmark Table Shows Uneven Strengths
Google presents Hydrogen as a substantial upgrade over Gemini 3.8 Flash. At Hydrogen’s default high setting, the company reports tasks completing 2.5 times faster, a Terminal-Bench 4.1 score of 57% versus Flash’s 18.2%, and more than a fivefold reduction in fabricated claims.
Its competitor comparisons show why model selection still needs workload-specific evidence.
Hydrogen leads these listed competitors on SWE-sweep, Agents’ Last Exam and HLE-Diamond. Claude Sonnet 5.5 leads Hydrogen on the cited Terminal-Bench and OSWorld evaluations.
That split is useful. Even within agentic work, repository repair, terminal execution and computer use reward different capabilities. A team selecting a model for a terminal-heavy workflow should give the Terminal-Bench result more weight than an unrelated reasoning score.
These percentages also use different benchmarks and scoring methods. A 71.6% score on one test cannot be read as directly better than 57% on another.
The defensible takeaway is narrower than a universal ranking: Google’s published results make Hydrogen a candidate for several coding and knowledge-work tasks, while documenting areas where another listed model performs better.
Google DeepMind states that Hydrogen is adapted from Gemini 4 Carbon. Its model card refers readers to Carbon’s documentation for architecture, training data, hardware and software details.
That establishes lineage. It does not disclose whether Google used distillation, pruning or another adaptation method, nor does it establish a parameter count. Claims about those details would go beyond the published information.
Google’s deployment strategy is clearer. It describes Carbon and Argon as models for difficult architectural decisions and long-horizon reasoning, while Hydrogen writes code, runs tests and verifies changes.
In Antigravity, Google’s new Orchestrator can switch between models such as Carbon and Hydrogen during the same session. Google reports that, on Terminal-Bench 4.0, Orchestrator approached Carbon’s task-success rate while consuming 24% less quota and finishing 19% faster on average.
Those figures describe a particular multi-model benchmark setup. Quota consumption is also a product-specific measure, so a 24% quota reduction should not be translated directly into a 24% API bill reduction.
The distinction between benchmark versions matters too: the Orchestrator claim uses Terminal-Bench 4.0, while Hydrogen’s standalone model-card comparison uses 4.1. They answer different questions and should not be combined into a single performance claim.
Why Worker Agents Are a Plausible Fit
Google also describes Carbon coordinating Hydrogen worker and verifier agents across isolated workspaces. Hydrogen can batch file-reading and search calls in parallel, according to the launch announcement.
The practical reasoning is sound: independent file inspections need not wait on one another, and separate workers can build or check different changes concurrently. Actual gains depend on tool latency, task dependencies and the amount of work that can safely happen in parallel.
My reading of the published evidence is that Hydrogen’s strongest deployment argument is selective execution. Keep expensive architectural decisions with the model that handles them reliably, then route bounded implementation and verification tasks to Hydrogen where evaluations support that choice.
Choose by Accepted Results, Not Token Price Alone
Hydrogen’s flat rates make budgeting easier, but the launch materials do not provide Carbon and Argon prices for a like-for-like API cost comparison. They also do not establish a universal routing threshold between the three models.
A practical evaluation should count completed, accepted tasks alongside tokens, elapsed time, retries and human corrections. A cheaper individual call can become an expensive workflow if it requires repeated repair.
For business and developer teams, the November 12 release offers two concrete options: use Hydrogen as a standalone model for recurring execution work, or evaluate it as a worker alongside Carbon and Argon. Start with tasks whose success can be checked clearly, such as passing tests, producing a valid structured response or making an isolated repository change.
Frequently Asked Questions
4 questions
1
When Did Gemini 4 Hydrogen Launch?
Google introduced it on November 12, 2026. The main announcement was updated on November 16.
2
What Does the Gemini 4 Hydrogen API Cost?
Google lists $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across the advertised context window.
3
Can Hydrogen Generate One Million Tokens?
The model card lists up to 64,000 synchronous output tokens and up to one million through background long decode, subject to its shared one-million-token context budget.
4
How Does Hydrogen Differ From Carbon and Argon?
Google positions Hydrogen for fast execution and everyday agentic work. Carbon and Argon target deeper, longer-horizon reasoning, and Hydrogen can work alongside them in multi-model workflows.