Google Releases Gemini 4 Hydrogen to Run Agent Tasks 19% Faster
Google rolls out Gemini 4 Hydrogen globally as its daily driver for agentic workflows, available free via auto model selection and across every product tier.
- Google launched Gemini 4 Hydrogen, its fast daily-driver model for agentic workflows.
- Rolling out globally across the Gemini app, Search, AI Studio, API, and Antigravity.
- Free users get Hydrogen automatically via auto model selection in the Gemini app.
- API priced flat at $1.50 input and $7.50 output per million tokens across 1M context.
- Scores 51 on Artificial Analysis Intelligence Index, edging out frontier model Argon at 50.
- Orchestrator pattern with Carbon cuts quota 24% and runs 19% faster on Terminal-Bench 4.0.
Google has launched Gemini 4 Hydrogen, a model designed for writing, tool use, interface prototyping, and other latency-sensitive agent work. It is available from launch through developer products, consumer apps, Search, and business offerings. For API teams, Hydrogen provides a worker model for routine steps inside multi-model agents, with published rates across a one-million-token context window.
Flash’s role gets a new name
Google has assigned Hydrogen the fast-model role that earlier generations labeled Flash. After launching Gemini 3.8 Flash in September, the company adopted element names across the Gemini 4 lineup. Each model now has a distinct role based on task complexity and execution speed.
| Model | Primary role |
|---|---|
| Hydrogen | Fast execution, writing, tool calls, and interface prototyping |
| Carbon | Long-horizon reasoning and coordination of agent workers |
| Argon | Frontier reasoning for difficult, long-running problems |
Three workloads define Hydrogen
- Writing and transformation: Google says Hydrogen follows instructions and tone constraints more closely, producing drafts, study guides, rewrites, tables, and visual summaries with fewer revisions.
- Connected agents: The model can break goals into steps, retrieve context from uploads and supported services such as Gmail, Google Drive, Airtable, and Dropbox, then select and invoke tools.
- Interface prototyping: Hydrogen is tuned to generate functional layouts and interactive mockups in fewer iterations.
Connected-app behavior depends on product support, user authorization, and each service’s permissions. API customers need to provide the relevant connectors, authentication, and tool schemas for agents running outside Google’s managed products.
Access spans six Google surfaces
| Product | Access |
|---|---|
| Google AI Studio | All tiers |
| Gemini API | All tiers |
| Gemini for business | Available at launch |
| Gemini app | Available at launch; free users receive Hydrogen through automatic model selection when Google considers it suitable for the prompt |
| Google Search | Available at launch |
| Google Antigravity | Available to paid Google AI Pro and Ultra subscribers |
Production teams should confirm the API model identifier, regional availability, rate limits, and release status shown in their Google console before pinning Hydrogen in deployed code.
Hydrogen edges Argon in one index
At the default high thinking level, Google reports that Hydrogen scores 51 on the Artificial Analysis Intelligence Index, compared with Argon’s 50. Hydrogen reaches 54 at the highest xhigh setting. Google also says it often completes tasks faster than other models in its class.
The Artificial Analysis index combines multiple evaluations into a general score. A thinking level controls how much inference effort the model applies before answering, and higher settings can increase response time and token use. Workload-specific tests remain necessary for evaluating instruction following, tool accuracy, tail latency, and total cost.
Long prompts keep base rates
Gemini 4 Hydrogen uses the same input and output rates across its full one-million-token context window. The context window represents the maximum tokenized content a request can hold, including the prompt, retrieved material, tool context, and generated response.
| Item | Published rate |
|---|---|
| Input tokens | $1.50 per million tokens |
| Output tokens | $7.50 per million tokens |
| Context window | Up to one million tokens |
Using those rates, a request with 100,000 input tokens and 5,000 output tokens costs $0.1875. Flat pricing across the context window gives retrieval-augmented generation systems, which attach retrieved documents to a prompt, a predictable base rate for large requests.
Carbon coordinates Hydrogen workers
Google’s Orchestrator pattern uses Carbon to plan and coordinate a task, then delegates routine steps to Hydrogen workers. On Terminal-Bench 4.0, an evaluation of agents completing command-line tasks, Google says the combination achieved a success rate near Carbon’s while consuming 24 percent less quota and finishing 19 percent faster.
Google’s phrase “near Carbon” leaves the exact quality gap unspecified, so the benchmark cannot establish the tradeoff for a particular application. Quota measures service capacity; billing savings depend on the mix of input, output, tool calls, and models used in production.
Put routine steps on Hydrogen
- Use Hydrogen for worker loops: Route drafting, extraction, tool calls, retrieval, and interface generation to the faster model.
- Escalate complex tasks: Send difficult planning and long-horizon reasoning to Carbon or Argon, then return bounded subtasks to Hydrogen.
- Test thinking levels separately: Measure success rate, latency, and token consumption at
highandxhighwith representative prompts. - Validate long-context quality: A one-million-token limit permits large requests, while retrieval quality and instruction adherence determine whether the extra context improves results.
- Track end-to-end usage: Compare total tokens, retries, tool failures, and quota consumption before applying Google’s benchmark savings to capacity or cost forecasts.
Comments
