Gemini 4 Hydrogen has four thinking levels: low, medium, high (the default) and a new xhigh. Google’s facts here come from its launch posts and API release notes. The advice on when to use each level, and all the prompts, are our own.
What Google says
- Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its 1 million token context.
highis the default. Google says it finishes tasks about 2.5x faster than Gemini 3.8 Flash at that setting.- Google reports a score of 51 on the Artificial Analysis Intelligence Index at
highand 54 atxhigh(Google’s own figures).
More thinking usually means more output tokens, so a higher level costs more per task even though the price per token is the same.
Which level for which job (our rule of thumb)
| Level | Use it for | Example |
|---|---|---|
low | Short, clear tasks where speed matters | Rewrite a title, extract fields, label a ticket |
medium | Everyday drafting and simple code edits | Summarise a document, write a small function |
high (default) | Multi-step work and debugging | Fix a failing test, plan a feature, review a pull request |
xhigh | Hard problems where a wrong answer is costly | Tricky bugs across files, careful analysis of long reports |
Start at the default. Drop a level when answers are already right and you want speed, and raise it only when a task keeps failing.
Prompts to try (our own)
Low: quick extraction
From the email below, return JSON with fields: sender_name, requested_date, product, urgency (low/medium/high). Use null if a field is missing. No other text.Medium: everyday drafting
Summarise this meeting transcript for someone who missed it. Give five bullet points, then a list of decisions and a list of open questions with owners.High: debugging
Here is a failing test and the function it tests. Explain the root cause in two sentences, propose the smallest fix, and list what else might break. Do not change unrelated code.xhigh: careful analysis
Compare these two vendor contracts section by section. Flag every difference in payment terms, notice periods and liability caps in a table, quote the clause numbers, and say where wording is ambiguous. I will check the flagged items with a professional.A quick cost test
- Pick three real tasks you do every week.
- Run each at
medium,highandxhighwith the same prompt. - Note whether the answer is right and how many output tokens it used.
- Keep the lowest level that gets it right.
Our prompt workflow explains how to score answers, and our prompt templates give you starting points.
When to switch model instead
If xhigh still struggles on long, multi-step work, try Google’s frontier models. Our Gemini 4 model guide compares Hydrogen, Argon and Carbon. For agents that act on your systems, add the safeguards in our agent permissions checklist.