FREE WEBINAR X NetApp
Quick Answer
Try building a playable game from a single prompt, recreating a webpage from a screenshot, digging through a huge PDF, searching a long video for one moment, researching the live web, planning a trip with Maps data, or handing it a multi-step task that needs tools. Google Gemini 3.7 Flash is a multimodal AI model built for coding, long-context reasoning, and agentic workflows, so the interesting part is watching it do more than chat.
Asking a new AI model to summarize a paragraph or draft an email doesn’t tell you much anymore, since nearly every model handles that fine. We wanted to see what makes Gemini 3.7 Flash worth using, so we tested it with code, screenshots, PDFs, huge context, video, live information, tools, and multi-step tasks.
A couple of notes first. The official name is Gemini 3.7 Flash, and some of what we cover needs Google AI Studio or the Gemini API rather than the consumer app. Also, Gemini 3.8 Flash exists and is newer, so we’re not calling 3.7 the latest here. Our piece on large language models is a decent refresher, and Google’s own launch post covers the coding and agentic angle.
1. Build a Playable Game from One Prompt
Coding is one of the areas Gemini 3.7 Flash is built for, and a browser game is a quick way to see if that holds up.
Ask for a game in HTML, CSS, and JavaScript, and be specific, think mechanics, scoring, enemies, levels, controls, and a restart button. Build a playable retro tower defense game is a fine starting prompt. What matters is whether the game runs and plays, not whether the code looks convincing.
Our roundup of AI coding tools is worth a look if you want to compare Gemini against other options.
2. Recreate a Website from a Screenshot
This one tests whether Gemini can connect what it sees to what it can build, a different skill from writing code on request.
Upload a screenshot of a landing page, dashboard, or ecommerce page and ask Gemini to recreate it, watching layout, spacing, typography, and responsive behavior. Then ask it to compare its own output against the screenshot and fix what doesn’t match.
That follow-up step usually reveals more than the first attempt. Google’s feature documentation for Gemini 3.7 has more on how it handles this.
3. Turn a Large PDF into an Interactive Dashboard
Summarizing a PDF is fine, but turning it into something usable is a better test.
Upload an annual report, a research study, or an industry report, and ask Gemini to pull out the data, spot trends, build charts, and put together a simple interactive dashboard.
If any number actually matters, double check them against the original first. Google covers document handling in its document understanding documentation.
4. Debug an Existing App
Writing new code is one kind of test. Fixing code that’s already broken is a more honest one.
Give Gemini the code, the error message, and what it’s supposed to do. Ask it to diagnose the root cause first without touching anything, then apply the smallest fix that solves the problem.
What you’re judging is the diagnosis, the fix, and whether anything else breaks along the way. This mirrors real debugging more than a fresh coding prompt does.
5. Analyze a Huge Collection of Documents
Gemini 3.7 Flash supports roughly a 1-million-token context window, so it can hold onto and reason across a genuinely large stack of information at once.
Feed it several reports, meeting notes, papers, or project documents together. Skip summarize all of these and ask a cross-document question instead, like find where these reports contradict each other.
That kind of question actually uses a long context instead of just filling it up. Our piece on AI token usage and optimization explains what a window like this means in practice.
6. Search through a Long Video
Gemini can watch and reason for video, and finding one specific moment in a long recording is a solid way to test that.
Try it on a podcast, interview, keynote, or lecture, and ask something specific rather than for a summary, like find every point where pricing comes up and give timestamps.
You can also have it flag where speakers disagree or where the topic shifts. Google’s video understanding documentation has the details.
7. Research the Live Web with Google Search
In supported workflows, Gemini can use Google Search grounding to pull in current information instead of relying only on what it already knows.
Ask it to research a company’s biggest announcements from the past month, and have it sort the answer into confirmed facts, company claims, independent reporting, and open questions. That structure says a lot about how it handles retrieval.
Details are in Google’s Search grounding documentation. This isn’t the same as retrieval-augmented generation, which our piece on vector databases for multimodal GenAI covers separately.
8. Plan a Trip with Google Maps Data
Google Maps grounding gives supported Gemini workflows access to real, location-aware data instead of generic travel advice.
Give it a handful of locations plus constraints, something like visit these six London attractions, minimize travel time, have lunch at 1 p.m., and finish near King’s Cross by 7 p.m.
The point is testing spatial reasoning, not just getting a list you could have found yourself. More detail is in Google’s Maps grounding documentation.
9. Complete a Multi-Step Task with Tools
An AI agent does more than answer a prompt, it plans intermediate steps and uses tools along the way to reach a goal.
Give Gemini a task with several stages, like researching products, collecting pricing, calculating annual costs, and recommending an option. This is where you see the gap between a model that generates text and one that executes toward an outcome.
Our piece on AI agents and agentic AI goes deeper, and Google’s tools documentation covers the technical side.
10. Let Gemini Navigate a Browser Interface
Gemini’s Computer Use capability can interpret screenshots and generate interface actions like clicks, typing, scrolling, and navigation in supported environments.
Stick to something low-risk, like navigating public pages or testing a workflow. This is still an advanced, developer-facing capability, and anything with real consequences needs a human keeping watch.
We’ve covered agentic AI infrastructure requirements separately, and Google’s Computer Use documentation has more.
11. Audit an Entire Website or Codebase
Most coding tests stick to a single file. Auditing an entire project shows how Gemini handles something bigger.
Ask it to audit performance, accessibility, SEO, code quality, and UI consistency across a whole site or codebase. Have it find issues, prioritize them, and lay out a plan before touching any code, then tell it to fix only the highest-impact items.
This test is less about one fix and more about whether Gemini holds priorities across a longer, messier workflow.
12. Compare Low, Medium, and High Thinking Modes
Gemini 3.7 Flash offers low, medium, and high thinking levels, which trade response speed against how much reasoning effort goes into the answer.
Run the same difficult prompt through all three settings and compare speed, accuracy, and usefulness. Worth flagging that more thinking isn’t automatically better, sometimes low is genuinely enough.
If cost matters too, our cost comparison of ChatGPT, Claude, Gemini, Grok, and DeepSeek is a useful companion read.
Gemini 3.7 Gets More Interesting When You Stop Using It Like a Chatbot
The pattern across all twelve is pretty clear. Gemini 3.7 Flash gets more interesting once you stop treating it like a chatbot and start giving it code, documents, screenshots, video, huge context, tools, and tasks with multiple steps.
If you’re not sure where to start, try these three experiments: turn a screenshot into a website, a PDF into a dashboard, and a prompt into a playable game. Those three alone tell you more than an hour of casual chatting would. And yes, Gemini 3.8 Flash is newer, but for 3.7 Flash specifically, everything above still holds.
Frequently Asked Questions
Google Gemini 3.7 Flash is a multimodal AI model built for coding, agentic workflows, and multi-step execution. It supports text, images, video, and document inputs alongside tool use.
Yes. Gemini 3.7 Flash remains available even though Google has released newer Gemini models since it launched.
No. Gemini 3.8 Flash is newer than Gemini 3.7 Flash and reflects Google’s most recent release in the line.
It handles coding, PDF analysis, image and video understanding, long-context analysis, Google Search grounding, Maps grounding, tool use, and agentic, multi-step workflows.
Yes. Coding and software engineering are among the major workloads Google designed Gemini 3.7 Flash to handle. Sections 1, 4, and 11 above cover a few ways to put that to the test yourself.
