Gemini 4 Leak? 10 Interactive 3D Tests Against GPT-6 Astra and Fable 5.1

Last weekend a mystery model that appeared on LMSYS Arena as gemini-3.8-flash was claimed to be an early checkpoint of Gemini 4 Pro. Arena (formerly Chatbot Arena) works by blind voting: two anonymous models answer, identities are revealed only after the vote; testing a new frontier model in the shadows would be neither first nor unique — the 2024 gpt2-chatbot episode wrote the playbook. Creator Dom turns the claim into a ten-test bake-off: one prompt, three rivals, a single-file HTML output.

Test one is a mechanical butterfly. Three.js (an open-source JavaScript library that draws 3D scenes in the browser via WebGL — imagine reading Lego instructions piece by piece and assembling the scene screw by screw) is asked to produce a single-file model: explode and assemble, isometric and plan views, wing beats and background color switches. The pipeline is roughly: 1) parse the description into geometry, 2) build wing, body and hinge meshes, 3) add an animation loop and camera controls, 4) pack a self-contained HTML file. In the video Gemini 4 looks smoothest, Astra splits autopilot and manual flight, and Fable 5.1 wins the cleanest explode-to-assemble transition.

The next family checks visual fidelity: a product rendered as SVG (scalable vector graphics — math-defined shapes that stay crisp at any zoom, not pixels) and whether the model can copy it faithfully. Gemini 4 Pro's Xbox rendering is near-indistinguishable; the older Gemini Flash 3.8 shows glitches, while Meta Muse and GPT-6 Astra land very close and Grok 4.6 leaves obvious distortions. An animated variant where every button works and the aesthetic stays minimal shows the model can wire interaction logic, not just shapes. In short: can the model look at a screenshot and turn 'what does this look like' into code?

The most ambitious demo is Voxel Pagoda — an interactive island made of voxels (3D cubes, the Minecraft block of spatial computing). Gemini 4's world lets you sweep through day, sunset, night and sakura dusk, tune the wind and orbit the island smoothly. Rebuilt with Astra the narrator calls it 'even better': teleport to Moon Bridge, feed koi, dim lanterns, ring the bell, trigger petal showers and walk with WHSD — all in one file. Fable 5.1 adds tea-house walks, a drone tour and 3D export. This is the real-world cousin of prototypes like Google's Project Genie: imagine, type, walk inside.

The video also checks drawing breadth: the same pen sketch is given to Gemini, Astra, Grok and Claude Opus, and the cheaper Claude Opus is judged most balanced this time. Then comes an interactive jet engine cross-section — turbine, combustor and compressor exploding and reassembling with close-ups. A pelican riding a bicycle recalls an old humorous SVG threshold, in the same lineage as Karpathy's 'show me the first paragraph of The Lord of the Rings as a Three.js world': stress tests that fuse conceptual grounding with humor. The Valley micro-robot, with neck extension, shoulder lifts, gripper moves and eye contact, follows the same pattern: one prompt, dozens of parameters, a live toy in the browser.

Two broader scenes close the loop. A 3D tram simulation looks neck-and-neck between Astra and Gemini, hard to separate on physics and motion. An Airbus H145 rendered for idle flight in a dark luxury showroom, amber sunset and cyber-aerial tones feels polished, and a racing-car game is generated in minutes. The speed matters: where a traditional pipeline of Blender plus code plus textures takes weeks, the model ships a playable prototype in minutes. The host's repeated 'comment for source code' underscores that these single-file artifacts are meant to be reproducible.

How solid is the leak? The most circulated scoreboard claims Gemini 4 beats Astra and Fable on wider context and lower price. Yet GadgetsFocus, checking on September 19, is crisp: Google confirms work on Gemini 4, but there is no public launch date, official model identifier or verified scores. Eyestech's teardown of the 04:23 UTC September 17 Arena tracker entry for gemini-3.8-flash and NokiPowerUser's analysis show strong correlation, but the chain still rests on anonymous votes plus screenshots plus single-file showmanship. On OpenAI's side the ledger is clear: GPT-6 Astra opened on September 3, 2026 to limited orgs, then to all paid users, billed as state-of-the-art on computer use, browsing, software engineering and with a claimed 98 percent on FrontierMath Tier 4.

My read: even if hyped, these clips signal a new threshold. Gemini 3 Pro shipped in November 2025 with new reasoning and agentic tooling; Astra and the Fable 5 family deepened the same lane in 2026. The contest is no longer 'who writes the better poem' but 'who builds the more consistent spatial model and makes it interactive in one file.' The winner on the triangle of price, context window and reasoning depth will be the one that truly cheapens everyday product prototyping. Yet a single-file show does not erase weeks of testing, security and maintenance in production — demos shine, products endure.