Google has released Gemini 4 Hydrogen, the fast, lower-cost member of its Gemini 4 family, and describes it as its "daily driver for agentic workflows." The model was announced on November 12, 2026, and Google says it finishes tasks about 2.5 times faster than Gemini 3.8 Flash while cutting fabricated claims by more than five times. It follows the larger Gemini 4 Carbon and Argon models.
Where Hydrogen Sits in the Gemini 4 Lineup
Google positions Hydrogen alongside its two frontier models, Gemini 4 Carbon and Gemini 4 Argon, which are aimed at deep, long-horizon reasoning. Hydrogen is the quicker option for everyday coding and agent tasks, and Google says it can run on its own or act as the high-efficiency execution engine in multi-model workflows with the other two.
That split reflects a common design pattern for agent systems: a large model plans, and a smaller, faster model carries out the many steps that follow. In Google Antigravity, an Orchestrator feature switches between Carbon and Hydrogen. Google says that on Terminal-Bench 4.0 this setup reached task success close to Carbon's while using about 24% less quota and finishing tasks roughly 19% faster on average.
The Numbers Google Reports
Most figures below come from Google's own announcement, and independent replication will take time. Treat them as vendor claims until outside testers publish results.
- Artificial Analysis Intelligence Index: 51 at high thinking, compared with 50 for Argon, rising to 54 at the new xhigh level.
- Speed: 2.5 times faster than Gemini 3.8 Flash at its high default setting.
- Terminal-Bench 4.1: 57%, against 18.2% for Gemini 3.8 Flash.
- Hallucination: fabricated claims cut by more than five times.
- Partner tests: in early tests on complex graphs, SQL and data-heavy enterprise workflows, Hydrogen ran 2x to 2.8x faster with 12% to 23% lower token usage than 3.8 Flash.
- Other results: 58.8% on HLE-Diamond, 52.3% on Vals Tax Agent Bench, where Google says it ranks first, and a fifth-place ranking on Agent Arena.
A caution on interpretation: a model that scores one point above a larger sibling on an aggregate index does not make it better for every job. Index scores blend many tasks, and Google's own framing keeps Carbon and Argon as the choice for the deepest reasoning.
Thinking Levels and Pricing
Hydrogen supports four configurable thinking levels: low, medium, high, which is the default, and a new xhigh tier. Higher levels spend more reasoning effort per request, which generally raises quality on hard problems at the cost of latency and tokens. Google's developer post points readers to the Gemini API "What's New" guide for the controls and does not prescribe which level suits which task, so teams will need to test their own workloads.
Pricing is $1.50 per million input tokens and $7.50 per million output tokens, flat across the full one-million-token context window. A flat rate matters for agent runs that carry long histories, because many providers charge more once a prompt crosses a length threshold.
Availability
Google is rolling Hydrogen out widely:
- Gemini app: available in the model dropdown, and free users get it through automatic model selection when it fits the prompt.
- Google Search: rolling out to all users globally.
- Developers: available in Google AI Studio and the Gemini API for all tiers, and in Android Studio.
- Google Antigravity: for paid Google AI Pro and Ultra subscribers.
- Business: through the API, the model dropdown, Antigravity and Gemini in Google Sheets for eligible subscriptions, where Google says outputs are generated more than 30% faster than with previous models.
Google describes tighter instruction following as a goal for the developer release, along with rapid iteration on repository coding, debugging, interactive web development and multimodal workflows.
Why It Matters
Frontier labs are competing as much on cost and speed as on peak capability, because agents make many model calls per task and small per-call savings compound. A model positioned as the cheap workhorse, with a premium model reserved for planning, is how many production agent systems are now structured.
If Google's claims hold up, Hydrogen narrows the gap between fast and frontier-class models on agentic coding benchmarks, and the 12% to 23% token reduction would translate directly into lower bills. The caveat is verification: the benchmark results are company-reported, the speed comparison is against Google's own prior model, and partner tests were early. Developers evaluating Hydrogen should run it on their own tasks, compare thinking levels, and measure cost per completed task rather than cost per token.
