Leaked: I Tested the Gemini 3.0 Pro 'X28' Checkpoint! The Results Will Rattle Competitors

This is the one we've been waiting for. I've spent weeks testing the rumored Gemini 3.0 Pro through hidden A/B tests, and the latest 'X28' checkpoint is a game-changer. Here's my first-hand report on what Google is building—and why its rivals should be worried.

This wasn't an official preview. It started with a grind. By monitoring network logs in AI Studio while running prompts on "Gemini 2.5 Pro," I started catching a different checkpoint ID: "2HTT." This was the first hint of Gemini 3.0 Pro.

It was rare—appearing maybe once every 40-50 prompts—but the quality jump was immediate and massive. Then, after a weird "nerfed" checkpoint (ECPT) that had me worried, a new one appeared: X28.

This is the one. "X28" feels like the real deal, a massive step above even the first "2HT" checkpoint. If this is what Google ships, it will redefine the developer landscape.

While the first "2HT" checkpoint was good, "X28" is on another level. It doesn't just complete tasks; it understands them.

My floor plan tests went from "good" to "this is real." X28 generated layouts with proper doors, sensible room flow, and even lighting controls for the furniture.

This is huge. Across multiple runs, X28 produced similar, coherent layouts. This near-deterministic behavior is a dream for developers who need reliable outputs.

The UI and visual output are stunning. It generated a network simulation with a slider and regeneration controls, nailing the typography and spacing on the first try. It's less "vibecode" and more "real designer."

The 3.js Pokéball was polished with colorful backgrounds. The Minecraft scene had rivers and clean lighting. It nailed a Rust CLI for image conversion. This isn't just a toy.

Across all the strong checkpoints, especially X28, you can feel it "thinking." There's a slightly slower first token, followed by a deliberate and steady stream of output. It's not just predicting, it's constructing.

The biggest hinge is tool calling. My initial tests with function calls were on point, with the model selecting the right tool immediately. If Google has trained this for real-world agentic workflows (like for Gemini CLI and Jules), it could be the most reliable model for building real applications.

Based on my real-world prompt suite—not just simple web-style demos—here's the breakdown:

It's not even a contest. X28 is clearly ahead in spatial reasoning and one-shot 3D scenes.

Gemini 3.0 Pro (X28) is at or above Opus for generative code polish and design taste.

It's highly competitive. On math and consistency, they trade blows. GPT-5 might slightly edge out on some physics simulations, but Gemini's consistency and incredible UI taste win on many practical, real-world tasks.

The recent Vertex AI leak (showing "Gemini 3.0 Pro" with an "11-2025" tag) confirms a release is imminent.

The only question is which version we get.

If we get the nerfed "ECPT" variant, it's still an upgrade over Sonnet but not the generational leap I've seen. But if Google has the confidence to ship what I tested with X28... then yes, this is that "3.5 to 4" moment all over again.

I'll be running my full benchmark suite the second this lands in public preview. This is going to be fun.

The discovery of the X28 checkpoint represents a significant milestone in the AI landscape. After weeks of rigorous testing, it's clear that this model variant delivers on the promise of genuinely next-generation AI capabilities. With its superior spatial reasoning, remarkable consistency, and impressive design capabilities, X28 could position Google's Gemini 3.0 Pro as a serious competitor in the AI market.

The timing of this discovery, coming just as OpenAI, Anthropic, and xAI are preparing their own next-generation models, suggests that we're entering an exciting phase of rapid advancement in AI technology. If Google does indeed release the X28 variant to the public, it could shift the competitive dynamics significantly in their favor.

Only time will tell which checkpoint Google ultimately chooses to deploy, but the potential demonstrated by X28 sets a high bar for what we can expect from the upcoming Gemini 3.0 Pro release.