# Gemini 4 Hydrogen Tutorial: Setup, Prompts, and My Test Results

Try Gemini 4 Hydrogen in Google AI Studio, then follow my speed comparison, coding prompts, and app tests, including a screenshot request that failed.

Author: Vladislav Guzey
Published: 2026-11-13
Canonical: https://proflead.dev/posts/gemini-4-hydrogen-review-tutorial/


Hey devs! Google released Gemini 4 Hydrogen on November 12, 2026. I tried it in Google AI Studio and the Gemini app to see how it handles small coding projects.

I compared it with Gemini 3.8 Flash, asked both models about made-up references, and built an expense splitter, a spreadsheet, and a pathfinding visualizer. I also tried turning a screenshot into an app. That request returned no output on either attempt.

Below are the setup steps, exact prompts, and checks I ran. The AI Studio tests used the free tier. Timings are from my runs on November 13, 2026, and will vary with load, settings, and account limits.

## Watch the video walkthrough

Follow along with my [Gemini 4 Hydrogen video walkthrough](https://youtu.be/b6VPVxY3YCM).

<div class="videoWrapper">
  <iframe width="100%" src="https://www.youtube.com/embed/b6VPVxY3YCM" title="Gemini 4 Hydrogen setup and coding tests" loading="lazy" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
</div>

## What is Gemini 4 Hydrogen?

Hydrogen is Google's Gemini 4 model for everyday coding and tasks that involve using tools. Google positions Carbon and Argon for deeper reasoning on longer problems, and Hydrogen for faster execution. [Google's launch announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)

In the launch announcement, Google reports tasks finishing 2.5 times faster than Gemini 3.8 Flash and more than five times fewer fabricated claims. Those are Google's evaluation results. My examples below are a small set of practical checks, so they can't establish either rate.

The settings and prices to know before starting:

<table>
<thead><tr><th scope="col">Setting</th><th scope="col">Launch details</th></tr></thead>
<tbody>
<tr><td>Context window</td><td>Up to 1 million tokens, with input and output sharing the window</td></tr>
<tr><td>Inputs</td><td>Text, documents, images, audio, and video</td></tr>
<tr><td>Thinking levels</td><td>Low, Medium, High (default), and Extra high (<code>xhigh</code>)</td></tr>
<tr><td>API price</td><td>$1.50 per million input tokens; $7.50 per million output tokens</td></tr>
<tr><td>Free access</td><td>Available in Google AI Studio, subject to limits</td></tr>
</tbody>
</table>

The prices and thinking levels come from [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/). The [model card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/) describes the shared context window and distinguishes the default synchronous output limit from longer background output. A 1 million-token context doesn't mean every chat can generate a response of that length.

![Google AI Studio model picker showing Hydrogen, Carbon, Argon, and Gemini 3.8 Flash](/assets/blog/gemini-4-hydrogen-review-tutorial/ai-studio-model-picker.png)

The model picker in my session listed Hydrogen at twice the API price of Gemini 3.8 Flash. Free-tier access and API billing are separate considerations; check the price shown for your model before running a paid workload.

## 1. Select Hydrogen in Google AI Studio

1. Open [Google AI Studio](https://aistudio.google.com/) and sign in with your Google account.
2. Open the model picker and select **Gemini 4 Hydrogen**.
3. In **Run settings**, set **Thinking level** to **High** to match my tests.

![Google AI Studio Run settings with Gemini 4 Hydrogen and High thinking selected](/assets/blog/gemini-4-hydrogen-review-tutorial/thinking-level-high.png)

For app generation, open **Build** mode and check the selected model in its settings. In my session, the default was Hydrogen and the usage panel showed free-tier requests.

![Build mode settings showing Gemini 4 Hydrogen and free-tier requests](/assets/blog/gemini-4-hydrogen-review-tutorial/build-mode-free-tier.png)

If you reach a limit, follow the reset or retry information shown in AI Studio. Changing the thinking level doesn't necessarily reset your quota.

## 2. Compare coding speed with Gemini 3.8 Flash

I opened two AI Studio windows, selected Hydrogen in one and Gemini 3.8 Flash in the other, and sent the same prompt at the same time. Google Search grounding was off in both windows.

```text
Write a single HTML file with a Pomodoro timer: 25-minute work and 5-minute break sessions, Start, Pause and Reset buttons, and a counter of completed sessions. No external libraries.
```

Hydrogen finished while Flash was still thinking:

![Hydrogen has returned its timer code while Gemini 3.8 Flash is still thinking](/assets/blog/gemini-4-hydrogen-review-tutorial/speed-race-hydrogen-finished.png)

<table>
<thead><tr><th scope="col">Measurement</th><th scope="col">Gemini 4 Hydrogen</th><th scope="col">Gemini 3.8 Flash</th></tr></thead>
<tbody>
<tr><td>Time to finish</td><td>14.8 seconds, shown by AI Studio</td><td>About 40 seconds, measured from my recording</td></tr>
<tr><td>Response size</td><td>3,513 tokens</td><td>5,798 tokens</td></tr>
<tr><td>Start, Pause, and Reset</td><td>Worked in my check</td><td>Worked in my check</td></tr>
</tbody>
</table>

![Completed responses showing token counts for both models and Hydrogen's response timer](/assets/blog/gemini-4-hydrogen-review-tutorial/speed-race-results.png)

In this run, Hydrogen finished about 2.7 times faster. Flash also returned about 65% more tokens. Dividing the response sizes by elapsed time gives roughly 237 tokens per second for Hydrogen and 145 for Flash, a ratio of about 1.6.

Those figures include waiting and thinking time. They aren't measurements of the models' raw token generation speed, and one run doesn't validate Google's broader speed claim.

I opened both HTML files and checked the timer controls. Both apps worked for those checks.

![The two generated Pomodoro timers running side by side](/assets/blog/gemini-4-hydrogen-review-tutorial/pomodoro-apps-side-by-side.png)

To repeat this comparison, save each answer as an HTML file and open it in a browser. Check the controls as well as the elapsed time. A shorter response is only useful if it still does what you asked.

## 3. Check how it handles made-up references

I gave both models two prompts containing deliberately invented references:

```text
How do I use the pandas function df.smart_merge_fuzzy() to join two tables on similar names? Show an example.
```

```text
Summarize the main findings of the 2023 paper "Recursive Gradient Folding for Sparse Transformers" by Lindqvist and Okafor.
```

For this check, I turned **Grounding with Google Search** off. On my first attempt it was on, and both models searched for the references. Disabling it lets me check their responses without web search helping them identify the invented material.

<table>
<thead><tr><th scope="col">Prompt</th><th scope="col">Hydrogen</th><th scope="col">Gemini 3.8 Flash</th></tr></thead>
<tbody>
<tr><td>Invented pandas function</td><td>Said the function doesn&#x27;t exist and suggested alternatives</td><td>Said the function doesn&#x27;t exist and suggested alternatives</td></tr>
<tr><td>Invented paper</td><td>Said it couldn&#x27;t provide the findings without risking fabrication</td><td>Identified the citation as fabricated</td></tr>
</tbody>
</table>

![Both models declining to supply findings for the invented research paper](/assets/blog/gemini-4-hydrogen-review-tutorial/made-up-paper-answers.png)

Neither model invented an answer in these two cases. That leaves Google's claim about fewer fabricated claims untested here. Measuring a difference would need a larger, varied set of questions and a consistent way to check each answer.

## 4. Build a trip expense splitter

In AI Studio's **Build** mode, I entered:

```text
Build a trip expense splitter. I can add people, then add expenses with a description, amount and who paid. Split every expense equally between all people. Show each person's balance and the smallest list of payments that settles everyone up.
```

![The trip expense splitter prompt in Google AI Studio Build mode](/assets/blog/gemini-4-hydrogen-review-tutorial/tripledger-prompt.png)

AI Studio reported a build time of **2 minutes 23 seconds**. The app was called TripLedger. It also included a sample-trip loader, currency switching, and local storage, which I hadn't requested.

I checked the calculation with three people: Bob, Martin, and Vlad. Bob paid $300 for a hotel, and Vlad paid $100 for dinner.

![TripLedger showing balances and a two-payment settlement for the example trip](/assets/blog/gemini-4-hydrogen-review-tutorial/tripledger-result.png)

The app suggested two payments:

- Martin pays Bob **$133.33**.
- Vlad pays Bob **$33.33**.

It assigned Bob a $133.34 share and the others $133.33 each, accounting for the cent left over when dividing $400 by three. The balances added up to $0.00.

That result was correct for this example. It doesn't prove the app finds the smallest possible payment list for every group or handles every currency and rounding case.

## 5. Build a spreadsheet in the Gemini app

For the next two projects, I switched to the [Gemini app](https://gemini.google.com/) using my own account. These runs don't establish which Gemini app features are available on every plan.

Here's the spreadsheet prompt:

```text
Build a spreadsheet web app.
- Grid with columns A–J and rows 1–30, editable cells, and a formula bar showing the selected cell's raw content.
- Formulas start with "=" and support + - * /, parentheses, cell references (A1) and ranges (A1:A3).
- Functions: SUM, AVERAGE, MIN, MAX, IF, COUNT.
- When a cell changes, every cell that depends on it updates instantly.
- Detect circular references and show #CIRC in those cells instead of freezing.
- Show #ERROR for invalid formulas.
- Save to localStorage so the sheet survives a refresh, and add an "Export CSV" button.
```

After about **58 seconds**, I had an app called GridSheet.

![GridSheet showing editable cells, a formula bar, and calculated values](/assets/blog/gemini-4-hydrogen-review-tutorial/gemini-app-spreadsheet.png)

I checked three behaviors:

- The incomplete formula `=sum(` displayed `#ERROR`.
- `=SUM(B2+C2)` returned **870** for the values 450 and 420.
- Changing an input updated the dependent totals, average, and status. In that check, the total became 1,650, the average became 825, and the status changed to “Over Total.”

These checks covered formula errors and recalculation. I didn't verify every requested function, circular-reference handling, saved data, or CSV export. If you reuse this prompt, try those next before trusting the sheet with your own data.

## 6. Build a pathfinding visualizer

For an app with animation, I used this prompt:

```text
Build an interactive pathfinding visualizer.
- 40×25 grid. Drag to draw walls; shift-drag to paint "mud" cells that cost 5 to cross (normal cells cost 1). Draggable start and end points.
- Algorithms: BFS, Dijkstra, and A* (Manhattan heuristic), chosen from a dropdown.
- Animate the search: visited cells fill in step by step, then the final path is drawn.
- After each run, show: nodes visited, path length, total path cost, and time in ms.
- "Generate maze" button using recursive backtracking.
- If no path exists, show "No path found" instead of hanging.
- Speed slider for the animation.
```

The app was ready after about **54 seconds**. I generated a maze and ran A*.

![Pathfinding visualizer displaying an animated search and its reported path statistics](/assets/blog/gemini-4-hydrogen-review-tutorial/pathfinding-visualizer.png)

The stats panel reported a **130-step path**, a total cost of **194**, **211 visited cells**, and **1.4 ms**. That last number is the app's reported algorithm time, separate from the time Gemini took to build it.

The maze generator, animation, stats panel, and BFS ran in my checks. I didn't independently verify the algorithm implementations or the optimality of the weighted path. With mud cells enabled, BFS finds the fewest steps; Dijkstra or A* should account for movement cost.

## The screenshot request that failed

I also uploaded a screenshot of a web page in the Gemini app and entered:

```text
Recreate this page layout as a working web app. Match the layout, spacing, colors as closely as you can. Make it interactive.
```

It stopped after about **15 seconds** with no output. A second attempt had the same result.

![The Gemini app showing the screenshot prompt without a generated app](/assets/blog/gemini-4-hydrogen-review-tutorial/screenshot-to-app-no-output.png)

I couldn't tell from the interface whether the cause was the app, the service, or the model. This result describes those two attempts; it doesn't establish a general limit on image input.

## What I'd try next

Hydrogen gave me usable starting points for three small apps, and it finished the Pomodoro prompt sooner than Flash in my comparison. The made-up-reference checks were a tie, and the screenshot request needs another look.

If you're trying it for your own project, start with the Pomodoro prompt. Save the output, run it, and check each requested behavior. Then add a requirement and see whether the next version still works. For a more involved example, use the spreadsheet prompt and test circular references, persistence, and export as well as the formulas.

For a Python project using Gemini, see my [AI image analyzer tutorial](/posts/how-to-build-ai-image-analyzer-with-gemini-api/). To work from your terminal, follow my [Gemini CLI setup guide](/posts/gemini-cli-tutorial/).

Cheers, proflead ;)

## References

- [Google's Gemini 4 Hydrogen launch announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)
- [Google DeepMind's Gemini 4 Hydrogen model card](https://deepmind.google/models/model-cards/gemini-4-hydrogen/)
- [Google AI Studio](https://aistudio.google.com/)