I Tested Gemini 4 Hydrogen: 4 Thinking Levels, 1 Task

I gave Gemini 4 Hydrogen the same coding task at four thinking levels: all four solved it without a mistake, but low finished in 14 seconds with 3,801 tokens while xhigh took 198 seconds and 54,078 tokens. On this task, more thinking did not change the result; it only made the run 14 times longer and its paid-API equivalent about 17 times more expensive.

This is a small test, not a benchmark. One task, one run per level, on November 14, 2026, in Google AI Studio. Below I describe what I did, what I measured and what this result does not show. For the model itself see What is Gemini 4 Hydrogen, and for independent measurements the Hydrogen vs Sonnet 5.5 vs GPT-6.1 Terra comparison.

Results

Thinking level	Time	Output tokens (incl. thinking)	Total tokens	Lines of code	Hidden tests	558 real posts	Paid-API equivalent
low	14.3 s	3,051	3,801	179	18 / 18	no difference	$0.024
medium	21.2 s	6,284	7,034	181	18 / 18	no difference	$0.048
high (default)	45.3 s	9,799	10,549	208	18 / 18	no difference	$0.075
xhigh	197.6 s	53,328	54,078	485	18 / 18	no difference	$0.401

The input was 750 tokens in every run. I calculated the cost column myself from the paid API price ($1.50 per million input tokens, $7.50 for output); the runs were made in AI Studio, so these are not billed amounts.

What was the task?

I asked for a dependency-free, single-file Node.js script that checks this site's blog posts. It reads the Markdown files in a folder and checks ten rules:

• whether the title exceeds 60 characters and seoTitle 65,
• whether the excerpt is between 120 and 160 characters and whether it ends with an ellipsis,
• whether the body contains a level-1 heading,
• whether a Turkish file contains an em dash,
• whether internal links carry the right language prefix, end with / and point to an existing post,
• whether the FAQ heading appears more than once.

I added details on purpose to make it harder: lengths are counted in Unicode code points (so an emoji is not counted twice), fenced code blocks and inline code are ignored, a ``` line inside a block opened with ~~~ does not close it, and #fragment or ?query in a link is not part of the path. The prompt came to 750 tokens and was identical, character for character, in all four runs.

How did I measure?

• Setup: Google AI Studio, model gemini-4-hydrogen, Grounding with Google Search and all tools off, no system prompt. A new chat for each level.
• Time and tokens: the values AI Studio displays. I timed the xhigh run with a timer in the page.
• Correctness, in two stages: first, 18 test files the model never saw (at least one per rule and per trap), with the expected result written by hand. Then the 558 real posts on this site: I compared the output of the model's script, file by file, with the output of a reference script I wrote myself. I also checked exit codes (1 if there are issues, 0 if none, 2 if the folder is missing).

All four scripts passed all four checks completely.

What did I see?

On this task, low was enough. For a well-defined job with rules spelled out one by one, the lowest level handled every trap correctly. I could not measure any accuracy gain from a higher level, because there were no errors left to measure.

Time and tokens do not grow linearly. From low to high, time grew about 3-fold and tokens 2.8-fold. From high to xhigh, time grew 4.4-fold and tokens 5.1-fold. The most expensive step is the last one.

xhigh wrote longer code, not more correct code. 485 lines, more than twice as many as high. The result was the same.

The default matters. Hydrogen's default in the API is high. On this task it gave the same result as low in 3 times the time and at 3 times the cost. Leaving the level at the default without testing it on your own work can mean paying for nothing.

Artificial Analysis's index points the other way, though: there Hydrogen scores 27 at low and 51 at high. So on hard, open-ended work the level changes the result a lot. The two findings do not contradict each other: my task was narrow and well specified.

What does this test not show?

• One task, one run. Rerunning the same prompt changes time and tokens; this test does not prove that low passes every time.
• The task turned out easy. Since all four levels scored full marks, it could not separate the levels by capability. On ambiguous, multi-step work or work touching a large codebase the picture may differ.
• AI Studio, not the API. Times come from the interface; network and quota conditions differ in the API.
• No rival. This post does not compare Hydrogen with any other model.

How to try this on your own work

1. Pick a task from your real work and first write the tests that will measure success for you.
2. Run the same prompt at low, medium and high; note the time and token count.
3. Use the lowest level that passes; keep xhigh for work where the lower levels fail.

How to set the level in the API is covered in the guide to migrating from Gemini 3.8 Flash to Hydrogen, and the free tier in Is Gemini 4 Hydrogen free.

Frequently Asked Questions

Which thinking level should I choose for Gemini 4 Hydrogen?

For a narrow task with clear rules, low was enough in this test. On hard, open-ended work, independent measurements show a large gap between high and low. The safest route is to start at the lowest level on your own task and raise it until your tests pass.

How slow is xhigh?

On this task it took 197.6 seconds; high took 45.3 seconds and low 14.3 seconds. As this was a single run, read it as an order of magnitude, not an exact ratio.

Are thinking tokens billed?

Yes. According to Google's pricing page, thinking tokens are billed as output tokens. In this test, most of the 53,328 output tokens of the xhigh run went to thinking.

Was the test run for free?

The runs were made in AI Studio. The dollar amounts in the table are what the same token counts would cost at the paid API price.

Can the results be reproduced?

The prompt, the 18 test files, the reference script, the model's four outputs and the results file are kept in this site's repository under scripts/experiments/hydrogen-dusunme-seviyeleri. Model output can change from run to run, so the numbers may not come out exactly the same.

Sources

• My own test, November 14, 2026, Google AI Studio (gemini-4-hydrogen): time, token and correctness results.
• Google, Gemini API pricing (checked November 14, 2026): the $1.50 and $7.50 prices used in the cost calculation, billing of thinking tokens.
• Google AI Studio docs, What's new in Gemini 4 Hydrogen: thinking levels and high as the default.
• Artificial Analysis, model leaderboard: Hydrogen's index scores by level (27 and 51).