Something unusual is happening on the LMSYS Chatbot Arena. AI researchers and developers testing blind prompts noticed that a model disguised under the alias gemini-3.8-flash was turning in performances far beyond what any lightweight tier could produce. It wasn’t long before tech trackers realized they were looking at an early preview checkpoint of Gemini 4 Pro.

The discovery quickly set social media ablaze. AI developer Hakm (@Hakm_x) was among the first to flag the suspicious performance boost to the broader community:

Shortly after, creator Bee (@thtbee_) posted a viral breakdown showcasing the model’s spatial visual capabilities and raw output density:

As reported in recent updates tracking NokiaPowerUser’s coverage on Gemini 4 Pro leaks, this isn’t the first time Google has used anonymous Arena deployments to stress-test next-generation architectures before an official key note. However, the sheer leap in code synthesis speed and 3D visual math has caught even seasoned engineers by surprise.

From 3D Airships to F1 Cars: Pushing Three.js to Its Limits

The most striking evidence of Gemini 4 Pro’s presence comes from complex spatial rendering tests using Three.js. Rather than outputting simplified code snippets or static 2D mockups, the model generated complete, production-ready WebGL scenes in a single pass.

Tech creator Harshith (@Harshith_x) shared a remarkable result where the model built a fully interactive atmospheric vehicle in record time:

AI researcher Lumina (@Lumina_x) ran a direct side-by-side comparison between the standard release model and the mysterious checkpoint, noting how rapidly the new candidate operates:

In even more demanding stress tests, developers asked the checkpoint to construct a modern Formula 1 race car from scratch. The prompt required a bespoke monocoque design, double-wishbone suspension with separate upper and lower A-arms at all four corners, PBR satin carbon fiber materials, realistic slick tires, and studio HDR lighting—all controllable via smooth camera orbiting.

Where previous generative models struggle with floating parts or broken geometry, this candidate authored clean, bug-free WebGL code that rendered accurately on the first attempt.

Feature CategoryStandard Production Flash TierStealth Arena Candidate (gemini-3.8-flash)
3D & WebGL CapabilitiesBasic box/cylinder meshes & partial scriptsBespoke geometry, working suspension A-arms & complex 3D scenes
Material RenderingSimple Phong / Lambert shadersCustom PBR materials (satin carbon fiber, clearcoat, rubber)
Generation SpeedFast for basic text tasksRenders full interactive 3D apps (F1 cars, airships) in 5–10 mins
Code CompletionRequires iterative bug fixesSingle-pass runnable projects with built-in orbit controls

What This Stealth Test Means for Google’s Next Big Rollout

Google’s decision to deploy flagship preview candidates under lightweight aliases allows their team to gather real-world telemetry without the immediate pressure of an official product announcement. The benchmarks coming out of these tests suggest that Gemini 4 Pro focuses heavily on long-horizon task execution, native multi-turn coding, and deep spatial reasoning.

For web developers, 3D artists, and software engineers, these early leaks offer a clear glimpse of where generative AI is heading. When a model can architect a complex, physically styled 3D application from a single text prompt in under ten minutes, the boundary between conceptual design and production deployment becomes almost seamless.

Add NPowerUser as a preferred source on Google News
Add NPowerUser as a preferred source on Google News

RELATED ARTICLESMORE FROM AUTHOR