Testing Gemini 3.5 Flash with My God-Level Prompts: Google Declares the End of the AI Chat Era!

This article is straight to the point, focusing on three core parts: Why is Gemini 3.5 Flash so strong? How to play with it for free? I am open-sourcing my own Gemini god-level prompts, with 8 major scenarios ready for use! I've also included a treasure egg at the end for everyone!

Table of Contents

1. Why is Gemini 3.5 Flash so strong?

Core Data: What does 289 tokens/s mean?

2. Strategic Judgment: Why Google is no longer obsessed with benchmarks

3. Hands-on Experience: Gemini 3.5 Flash free entry + gameplay

4. Open-sourcing my own god-level Gemini Prompts:

1. KERNEL-X Ultimate Dynamic Expert System
2. HE COMPILER IDENTITY (Full-stack Compiler Persona)
3. The only Vibe Coding prompt you need
4. Naval Ravikant Clone (Decision-making tool, Philosophy/Life Prompt)
5. AI Workflow Expert System Prompt
6. Cognitive Mining Prompt
7. Deep Research Report Prompt
8. God-level Reflection Prompt

5. Final Thoughts: What is the endgame of this AI race?

1. Core Data: What does 289 tokens/s mean?

• Speed: 289 tokens/s (2.1x faster than 3.1 Pro, 4x faster than GPT-4o)
• Intelligence: AA Intelligence Index 55 (3.1 Pro is 57)
• Hallucinations: Dropped by 31 percentage points, a leap in reliability
• Agent Capability: GDPval-AA Elo 1656 (3.1 Pro 1314), MCP multi-step workflow 83.6%
• In short: Flash is the first to max out speed, intelligence, and reliability simultaneously.

Let me explain what these numbers mean in plain language:

An output speed of 289 tokens/s: You hit enter, and before you can even switch back to check a message on WeChat, 3000 words are finished. While 3.1 Pro is still contemplating life, GPT-4o and Claude Opus are left in the dust by 4x.

Intelligence Index 55: Only 2 points behind the flagship 3.1 Pro from four months ago.

The 31 percentage point drop in hallucination rate is the most insane part. I used to be nervous using Flash for Agents because it would make things up.

Now, when I let it run multi-step workflows, it actively stops to ask when it's unsure. This is completely different from the old way of just making things up; it's a qualitative change.

2. Strategic Judgment

Why is Google no longer obsessed with benchmarks?

To be honest, before Flash came out, I would hesitate for a long time every time I started a new conversation.

Using Pro, I could brew a cup of coffee while waiting for it to finish thinking.

Using Flash, I was afraid it would hallucinate at a critical step, wasting everything that came before.

For three years, we've had to choose between speed and accuracy, enduring this since GPT-3.5.

Flash has finally solved both issues at once.

3. Free Hands-on Experience

I use ZenMux to test Flash, which is the aggregation platform I always use to run models.

I saw it went live for free with 3.5 Flash last night and have been testing it all day. You can use it immediately without an API or a VPN.

It has a multi-model PK (battle) function where the same prompt can be sent to Flash, Claude, and GPT simultaneously to see which performs best.

Some of the 8 prompts I'm sharing were refined using this PK mode.

But I should also mention the boundaries so people don't get the wrong idea.

Flash isn't omnipotent. When I ask it to write long articles (3000+ words), the quality visibly drops in the second half, and the logic starts to drift. For deep long-form writing, I still go back to Pro. After all, Flash is a speed specialist, not an endurance runner.

My current division of labor is: Agents for workflows, batch processing, and instant dialogue—all to Flash. Deep thinking, long-form creation, and complex reasoning—still Pro.

After running this setup for a day, the feeling is:

The hesitation of the chat box era disappeared today.

4. Open-sourcing my own god-level Gemini Prompts

ZenMux has opened free trials, no need to mess with APIs or VPNs, just open and run: https://zenmux.ai/invite/DIX0WC

Use these battle-tested prompts to experience the power of Gemini 3.5 Flash to the fullest.

KERNEL-X Ultimate Dynamic Expert System (My trump card prompt, iterated 10 times)

Full-stack Reality Compiler Prompt—Compiles vague real-world needs into logically rigorous, conceptually original, and highly transparent "high-dimensional intellectual assets."

This is the only Vibe Coding prompt you need:

It takes ten years to understand Naval; I used 1 prompt to encapsulate it into a wealth and clarity operating system.

Complex Problem Analysis Prompt

God-level Reflection Prompt

AI Workflow Expert System Prompt

Deep Research Report Prompt

5. Final Thoughts: What is the endgame of this AI race?

As I write this, I want to talk about a bigger issue.

What caught my attention with the Gemini 3.5 Flash release wasn't the 289 tokens/s or the 31% drop in hallucinations. Those numbers will be overshadowed by new ones in six months.

What caught my attention is that Google's logic for taking action has changed.

The old logic was: My model is smarter than yours, so you should use mine.

OpenAI and Anthropic are still playing that game—GPT-5 rumors, Claude's reasoning depth, various benchmark arms races—it's a battle of IQ.

Google's logic this time is: I'm not competing to see who is smarter; I'm competing to see who can make AI actually run.

What does 289 tokens/s mean? It means Agents have gone from something "theoretically useful" to something that can be "plugged into a workflow this afternoon." A person sitting at a computer, finishing a dozen steps in five minutes, calling tools, and correcting their own mistakes—this used to only exist in demo videos, but now it's real.

This is the prototype of the endgame.

The endgame of chat is Agents. The endgame of Agents is infrastructure.

What does that mean? When AI becomes fast enough, reliable enough, and cheap enough—it is no longer a "product."

It is like water, electricity, or gas. You don't wake up every morning thinking "should I use some electricity today?" You just flip the switch.

The future of AI is the same: you will no longer hesitate over which model to use, which window to switch to, or wait seconds for it to think—you just take one action, and it has already finished the job.

What Google did this time was push the "flip the switch" action a huge step forward.

So what are OpenAI and Anthropic doing?

To be honest, it's not that they don't see this trend; it's that their ships have too much inertia.

Flagship models have high inference costs, long inference times, and business models tied to the "smarter model" line—shifting to speed-oriented products would disrupt their own flagship pricing systems. This isn't a technical problem; it's a commercial shackle.

So Flash's opponent this time isn't actually GPT-4o or Claude Opus; it's Google's own past—the Google that only knew how to fight with benchmarks.

Google finally stopped using "intelligence" to prove itself and started using "speed" to define the rules of the game.

Finally, a word on how this affects you and me.

For Agents, the threshold has shifted from "technical ability" to "imagination."

Previously, you had to know how to build LangChain, write tool-calling chains, and stumble through various framework pitfalls to get a barely usable Agent running. Now the threshold has dropped to—can you think through a task and spend five minutes explaining it clearly to the model.

Ultimately, the endgame of the AI race isn't about who beats whom; it's that one day you look back and realize the things that used to take half a day now only require one sentence.

And that day is closer than anyone thinks.

Egg Moment

Sharing a treasure prompt website that pairs perfectly with the powerful Gemini 3.5, it can definitely solve most of your problems and confusion in work, life, and study! http://prompts.chat — currently the strongest open-source free AI Prompt community! Thousands of high-quality prompts categorized by industry/scenario: - AI Agents (what I use most to really take off) - Business analysis, writing, code, marketing, image generation... Search keywords to get the best template instantly, just copy and paste!

Features:

• Completely free + open source
• Community-driven, with new god-level prompts every day
• Supports building private libraries (a boon for teams/privacy advocates)
• 158k stars on GitHub, praised by big names at OpenAI and Hugging Face You no longer have to blindly fumble with prompts; your efficiency will skyrocket. Highly recommended!