On 12 November 2026 Google introduced Gemini 4 Hydrogen, the fast and inexpensive model in the Gemini 4 family. It is meant to handle everyday coding and agentic workflows, while the larger Gemini 4 Carbon and Gemini 4 Argon remain the models for hard, long-horizon reasoning. Hydrogen is available now in the Gemini API, Google AI Studio, the Gemini app and Google Search. It costs $1.50 per million input tokens and $7.50 per million output tokens. Most of the performance figures come from Google itself, which is worth keeping in mind.
What Gemini 4 Hydrogen is
In the announcement, Koray Kavukcuoglu, Google’s Chief AI Architect, describes Hydrogen as a model for “rapid execution”. The division of labour is clear: Carbon and Argon plan, make architectural calls and track down subtle bugs. Hydrogen writes code, runs tests and verifies changes. Google also positions it as the direct successor to Gemini 3.8 Flash.
We covered the two larger siblings in our pieces on Gemini 4 Carbon and the general availability of Gemini 4 Argon. Hydrogen is the third model in the line-up and the first that Google explicitly ships to every tier of user, including free accounts.
Technical specifications
The developer documentation lists the key facts:
- Model ID:
gemini-4-hydrogen, available through GenerateContent and the Interactions API. - Context window: one shared window of 1,048,576 tokens for input, thinking and output, as with Carbon and Argon.
- Output: up to 65,536 tokens per request. Longer responses, up to the limit of the context window, are possible through background execution or client-side continuation.
- Thinking levels:
low,medium,high(default) and the newxhigh. The model rejectsminimalwith an error. - Free tier: Hydrogen can be used on the free tier of the Gemini API and in AI Studio, subject to rate limits.
For developers moving from Gemini 3.8 Flash, the migration section matters more than any benchmark. Google lists several changes that can break existing code or make it more expensive. The server ignores sampling parameters such as temperature, top_p and top_k. Explicit context caching (cachedContents) is not supported; implicit caching kicks in automatically for prompts of at least 4,096 tokens instead. And because Hydrogen thinks at high by default, an unchanged call may use more tokens than it did on Flash, which defaulted to medium.
Pricing at a glance
Google charges a flat price for Hydrogen across the entire context window:
- Input: $1.50 per million tokens
- Cached input (implicit caching): $0.075 per million tokens, a 95 percent discount
- Output including thinking tokens: $7.50 per million tokens
The fact that thinking tokens are billed as regular output is the key point for budgeting. According to Google, xhigh gives the best results but also generates the most tokens. Google itself recommends starting with high, running all four levels against your own evaluations and then picking the lowest level that meets your quality bar. That is sensible advice, but it assumes you measure.
Benchmarks: what Google reports
The following figures come from Google’s announcement. Independent verification exists only for the individual leaderboards that Google itself cites.
- Artificial Analysis Intelligence Index: 51 points at the default
highlevel, 54 atxhigh. Google puts Gemini 4 Argon at 50. - Terminal-Bench 4.1: 57 percent versus 18.2 percent for Gemini 3.8 Flash.
- Speed: tasks are said to finish 2.5 times faster than with 3.8 Flash at the default level.
- Fabricated claims: Google reports more than five times fewer than with 3.8 Flash.
- Agent Arena: rank 5 on arena.ai’s live leaderboard.
- Other results: HLE-Diamond 58.8 percent, Vals Tax Agent Bench 52.3 percent, Vals Legal Research Bench 55.8 percent, Multimodal Knowledge Agent 53.4 percent, Agents’ Last Exam v1.1 71.6 percent.
In early testing, partners reportedly saw runs 2 to 2.8 times faster and token usage 12 to 23 percent lower on workloads involving graphs, SQL and enterprise data, compared with 3.8 Flash. Google does not say who those partners are or how they measured. Claims like these are an indication, not proof.
The comparison with Argon stands out: a smaller, cheaper model edging ahead of the bigger one on the Intelligence Index sounds contradictory at first. But an index built from many individual tests does not show how well a model holds up on very long, branching tasks. That is exactly where Google still recommends Carbon and Argon. Anyone deploying Hydrogen should test it on their own realistic tasks, not on the index.
Orchestrator: several models in one session
Alongside Hydrogen, Google is adding an “Orchestrator” to its Antigravity development environment. It switches between models within a single session, for example Carbon for planning and Hydrogen for execution. On Terminal-Bench 4.0 the Orchestrator is said to come close to Carbon’s success rate while using 24 percent less quota on average and finishing 19 percent faster. For parallel agents, Carbon can act as coordinator while Hydrogen powers worker and verifier agents in isolated workspaces.
The same pattern can be built through the API: according to the documentation, a multi-turn session can route planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen using previous_interaction_id. The managed “Antigravity” agent now uses Hydrogen by default. The idea of reserving expensive models for the hard steps is familiar from other vendors too; Anthropic, for instance, still recommends Opus over the new Claude Fable 5.2 for much everyday work.
Safety and reliability
Google also gives figures for behaviour in agentic scenarios. Hydrogen is said to reduce overreach in agentic tasks by up to 81 percent and to cut misaligned actions and task shortcutting across tool-use and coding evaluations by a factor of 12. When faced with blocked or impossible tasks, it is reportedly up to two times less likely to bypass permission boundaries. Against indirect prompt injection, it is said to be more than 2.3 times more resilient on the Gray Swan IPI benchmark.
These are internal measurements or ones chosen by Google. They indicate a direction but do not replace your own safeguards. Anyone running agents with access to files, repositories or production systems should keep permissions tight and review outputs, regardless of how well a model scores in Google’s tests.
Who should switch
- Gemini 3.8 Flash users: Google recommends Hydrogen as the direct successor. Work through the migration checklist first: model ID, thinking level, caching, sampling parameters, tool declarations.
- Teams on Carbon or Argon: Hydrogen can take over execution steps and lower costs. According to Google, the request schema stays the same.
- High-volume, cost-sensitive applications: for pure classification and routing, Google still names Gemini 3.5 Flash-Lite as the cheapest option.
Conclusion
Gemini 4 Hydrogen is less a breakthrough than a sensible gap that Google is now filling: a fast working model for agents, designed to cooperate with the large Gemini 4 models. The price of $1.50 and $7.50 per million tokens is clearly stated, and the migration notes are concrete. The impressive benchmark jumps, on the other hand, almost all come from Google itself. If you plan to switch, test on your own tasks, choose the thinking level deliberately and keep an eye on the token bill.




