gemini-4-hydrogen.
The 1,048,576-token window is shared by input, reasoning, and the reply. A synchronous request stops at 65,536 tokens. Generation can continue up to the window minus the input, with background execution or client-side continuation. The provider accepts text, image, audio, video, and PDF. Output is text. This page does not claim that the Tess picker exposes audio or video.
When to use Hydrogen
Use Hydrogen for everyday code, for an agent that has to keep going, and for a document that should come back as text you can review. Move to Argon or Carbon when the task needs long reasoning or more checking.What Google measured
The 12 November announcement compares Hydrogen mainly with Gemini 3.8 Flash and with models in the same kind of work.- On the Artificial Analysis Intelligence Index, default high thinking scores 51, ahead of Argon at 50. At
xhighit reaches 54. - Against Gemini 3.8 Flash, Google says default high finishes the task 2.5 times faster, lifts Terminal-Bench 4.1 from 18.2% to 57%, and cuts fabricated claims by more than 5 times by flagging missing context instead of filling it in.
- On HLE-Diamond the announcement cites 58.8%. On Artificial Analysis that figure is the
xhighscore. High is 54.9%, and Argon high is 56.7%. - Agents’ Last Exam v1.1 is 71.6%. Multimodal Knowledge Agent is 53.4%. Agent Arena rank is 5th.
- Partners cited by Google saw 2 to 2.8 times the speed and 12% to 23% fewer tokens than 3.8 Flash, on graphs, SQL, and data workflows.
Orchestrator, which switches between Carbon and Hydrogen in one session, is a Google Antigravity feature. It is not a Tess control.
What Artificial Analysis measured
The published comparison splits the four levels. The times below are for the high variant, which is thegemini-4-hydrogen default, and for xhigh.
On Terminal-Bench 4.0, high scores 54.0% and Argon high scores 57.1%. On AutomationBench-AA, 74.4% against Argon high at 77.5%.
Where it does not lead
- Google still points Carbon and Argon at long reasoning. Hydrogen is the fast model, not a replacement for Carbon.
- On Terminal-Bench 4.0 and AutomationBench-AA, high trails Argon high.
xhighraises the score and the wait. On the index, time to first answer moves from 16 seconds to 44.temperature,top_p, andtop_kare ignored inside the allowed range. Outside that range the API returns 400. Theminimallevel returns an error.- The model card knowledge cutoff is March 2026.
Where it creates more value
Everyday code
The change fits in one pass: fix the bug, adjust the test, and leave the neighboring file alone. Hydrogen reads and searches in parallel and keeps going when a tool comes back empty. Expected result: the diff for this step, the test that passed, and what was left out.A document that becomes text
The material mixes a PDF, a table, and a chart. The answer needs a source, and it needs to say when the file does not contain the figure. Expected result: the requested text, the source of each number, and the gap when the document does not say.An agent that must not invent
The task has a tool, a search, and a hole in the context. The announcement says Hydrogen flags what is missing instead of filling it in, with more than 5 times fewer fabricated claims than 3.8 Flash. Expected result: what the tool returned, what was missing, and the question that still needs an answer.How to apply it in Tess
1
Start on Hydrogen
Use Hydrogen for everyday code, the agent, and the document. Choose Carbon when high still does not finish.
2
Define what done means
State the deliverable, what must not change, and the check that ends the task. Ask for the file or the diff, not a plan of what it would do later.
3
Set tools and approvals
Enable only the tools you need. Require confirmation before sending, publishing, or changing data.
4
Pick the thinking level
LOW for a queue and for classification. MEDIUM for chat that should answer quickly. HIGH, the default, for code, tools, and documents. XHIGH only when high fell short, knowing that the wait and the spend go up.
5
Sample before you repeat
Run one full case, check the result, and only then repeat the flow.
Three playbooks to start
1. A fix that fits in one pass
Input: the file, the expected behavior, and the test for this step.2. A document with a source
Input: the PDF or the spreadsheet, the question, and the answer format.3. An agent that stops at the hole
Input: the task, the enabled tools, and what counts as an answer.Limits and good practice
- Default thinking on
gemini-4-hydrogenis HIGH. It spends more tokens than a MEDIUM default. - A 1,048,576-token window does not replace selection. The reply can use only what remains after the input.
- Explicit cache does not exist. Put the stable prefix at the start of the prompt. Implicit cache looks at prefixes from 4,096 tokens.
- A synchronous request still stops at 65,536 tokens. Going past that needs background execution or continuation.
- Keep a human approval for financial, legal, clinical, or other high-impact decisions.

