🚨 AI Leaks | Token Gremlin

Among the top 1% of ChatGPT power users 🤖
I track AI leaks so you don’t have to 🔥
Frontier models and future tech 🦾

LEAKS:

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

This is wonderful. Just imagine how many educational apps we could create across different fields. Children would have so many new ways to learn.

[How To Prompt @HowToPrompt__]

A dad vibe coded a language app for his daughter where you just point your camera at anything and learn the word for it..

and it won an apple design award.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

Is this true? If so, GPT-5.6 Pro is beyond what we were imagining in the other posts. You can literally create incredible applications across many areas.

[Zaur Ibrahimov @zaur_ibra]

🚨 GPT 5.6 Pro is insane with @threejs

https://t.co/3u0TFw0YOy

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

Early hands-on tests of “Bidi 1”, OpenAI’s upcoming bidirectional voice model, are starting to surface.

The model is expected to land in ChatGPT, with a possible path toward Codex later as well.

What makes it interesting:

It can keep listening while speaking, instead of treating voice like a strict turn-by-turn exchange.
It can jump between tasks mid-thought without losing the thread.
It handles interruptions, pauses, and conversational overlap far more naturally.
It seems noticeably stronger at preserving spoken context as the conversation unfolds.

There is still a limit on how long it can talk continuously, which is not surprising. Even so, in testing, it reportedly counted up to 23 without needing to stop.

Bidi 1 is not publicly available yet, but judging by the recent signs and preparation work, it feels very close.

Cc: @testingcatalog

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

GPT-5.6 Pro Early Coding Output.

This model just generated a fully procedural, animated, controllable spider from scratch.

Eight-legged motion. Real-time control. Procedural behavior.
The kind of coding output that makes you stop and go: “okay… this is getting serious.”

Cc: @JaydenDavisNC

[jason @jxnlco]

Post your best codex billboard

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

@jxnlco

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

@thsottiaux GPT-5.6, I think.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

One of the people responsible for Codex said that a new era is unfolding right in front of us, and many still have not realized it.

He suggested that users should simply speak for 10 uninterrupted minutes, dumping every strange idea, constraint, nuance, edge case, intuition, and half-formed thought they have about what they want to build.

The model has become so structurally complex and technically advanced that it can infer intent from messy, high-dimensional human input and produce something remarkably close to what the user actually had in mind.

[Guinness Chen @guinnesschen]

Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

This is honestly wild.

DevDude is building WizardGenie, a Roblox vibe coding plugin for his AI game engine. It supports tools like GPT Codex, Claude Code, Cursor, Grok, Minimax, Kimi, BYOK models, and even local GPU models.

It can pull assets from the Roblox asset store, write code, create environments, and help build game worlds directly inside Roblox.

Still in alpha, but already looking very promising for AI assisted game development.

Cc: @oldgamesnob

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

OpenAI argues that the technical and computational capability of AI models for hacking already exists, and that hiding it from legitimate defenders worsens the imbalance. That is why they want to make it available to verified individuals under governance.

Anthropic argues that the capability exists but is far too dangerous, poses a national security risk, and must be contained.

I agree much more with OpenAI, because malicious actors exist, already have the technology, and will attack. Taking defensive weapons away from the public only makes everything worse.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

So far, it has only been possible to identify the GPT-Bidi-1 voice model in the latest ChatGPT update. Some elements remain encrypted, which leads me to believe that GPT-5.6 may be coming, but that we might not be getting the SuperApp yet.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

The new OpenAI voice model is already showing up in the web interface.

The launch of GPT-Bidi-1 is imminent, we believe it will happen later this week.

Cc: @testingcatalog

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

GPT 5.6 Pro’s SVG generation is on another level.

The detail, structure, and overall visual consistency here are seriously impressive. It’s wild to think back to the moment when Gemini’s results felt groundbreaking. The progress since then has been massive, and this output makes that very clear.

Prompt: "A hyperealistic image of Steve Jobs, from Apple, holding up the iPhone. The Apple logo in the background should be slightly blurred, but still noticeable. The expression on Steve Jobs should be exaggerated, showcasing both excitement, thrill, and joy. Extreme detail. Go wild and be accurate"

Cc: @JaydenDavisNC

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

Another test, and honestly… this one is pretty adorable.

GPT-5.6 Pro: a tiny 3D voxel pelican riding a bike.

Cc: @JaydenDavisNC

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

What makes this Codex billboard interesting is that it does not feel like a typical AI product ad.

It does not show a sterile lab, a corporate dashboard, or some abstract promise of “productivity.” It shows something warmer and stranger: a person building in a private room, a desk, a screen, a dog nearby, and the city alive outside.

The contrast is the message.

One billboard feels like construction: a workshop, tools, assembly, the physical act of making something. The other feels domestic: home, focus, warmth, quiet creative work. Together, they suggest that Codex is not only for hardcore programmers staring at terminals. It is for people using code as a way to materialize ideas.

Even the location matters. Christopher Street and the West Village carry this sense of art, counterculture, independence, nightlife, and lived urban texture. Placing Codex there makes it feel less like enterprise software and more like a creative tool for independent builders.

And then there is the street itself: the bar, the lights, Village Cigars, the crowd, the old buildings. A very human, analog city. Above it, Codex appears almost like a digital layer floating over the physical world.

That is the visual sentence here:

the future is not hidden inside a white laboratory.

It is attached to the city, above the bar, inside the bedroom, on the desk, in the middle of human chaos, quietly turning intention into software.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

Look at these 3D renders that Fugu Ultra produced. Sakana AI came in strong.

This means we are going to see a clash between very powerful AIs in the coming weeks, because OpenAI is not just standing still and watching.

Cc: @omarsar0

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

@l_mejiaC I hate waiting. @OpenAI, give me GPT-6 now. @Anthropic, give me Fable 6 now!!!!!!!!!!!

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

GPT 5.6 Pro has just generated a complete 3D cinematic.

The characters, environments, animations, camera movement, and action choreography were all produced by AI, with no external assets.

The result is genuinely impressive, while also revealing some of the model’s current limitations. Across nearly 10 seconds, the ninjas were mostly just trading sword strikes, so that section was removed to keep the sequence tighter and more visually effective.

That said, the falling scene followed by the kunai throw stands out as particularly strong. It shows how, with precise direction and well-structured prompting, the model can produce cinematic shots that feel dynamic, stylish, and remarkably polished.

Cc: @mirochill

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

Prompt:

Build a polished self-contained browser 3D FPS game called ЛЕСНОЙ РУБЕЖ / FOREST STRIKE — ТАКТИЧЕСКИЙ ШУТЕР, matching a stylized low-poly forest tactical shooter prototype. Use Three.js/WebGL with procedural geometry only. Start with a dark Russian main menu: big title ЛЕСНОЙ РУБЕЖ, glowing lime button В БОЙ, controls panel in Russian, and footer FOREST STRIKE THREE.JS PROCEDURAL. After clicking play, spawn the player in a low-poly forest outpost with pine trees, leafy trees, birches, grass, bushes, rocks, sandbags, crates, barrels, log cabins, wooden watchtowers, fences, a campfire with smoke/fire particles, distant pale green mountains, blue sky, bright sun glare, and atmospheric fog.
Implement full first-person gameplay: WASD movement, mouse look with pointer lock, shift sprint, space jump, C crouch, LMB shoot, RMB aim down sights, R reload, mouse wheel and keys 1–4 weapon switching, F weapon inspect. Add first-person low-poly arms with camouflage sleeves and a stylized black AK-like rifle with optic, pistol, knife, and F-1 grenade. Rifle must show weapon sway, recoil, muzzle flash, bullet tracers, reload animation, 30-round magazine, reserve ammo, and centered circular optic when aiming. Grenade must visibly throw through the air, spin, explode after a delay, create smoke/particle burst, and damage enemies.
Use a minimal Russian HUD: top-left translucent kill counter УНИЧТОЖЕНО: 0, bottom-left ЗДОРОВЬЕ green health bar, bottom-center weapon slots АВТОМАТ / ПИСТОЛЕТ / НОЖ / ГРАНАТА with lime active highlight, bottom-right weapon name/ammo/fire mode such as AK-105 “КЕДР”, 30 / 150, АВТОМАТИЧЕСКИЙ. Add small white crosshair. Display the objective text ЗАЧИСТИТЕ ЛЕС ОТ ПРОТИВНИКА after starting.
Add simple low-poly enemy soldiers placed near trees, towers, cabins, and cover. They patrol, detect the player, move or shoot, take damage, and disappear/collapse when defeated. No gore. Increase kill counter on enemy death. Add impact particles, enemy hit flashes, damage vignette, and simple enemy muzzle flashes. The final result must look like a playable low-poly forest FPS prototype, not a static scene, not a generic shooter, not sci-fi, not horror, not zombies, not third-person, not pixel art. Prioritize matching the exact visual composition: forest clearing, bright sun, foggy mountains, watchtower, campfire, crates/barrels, AK rifle at bottom-right, ADS optic centered, Russian HUD with lime-green accents.

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

GPT-5.6 Pro vs Fable 5

Both created a 3D FPS forest shooter from a single prompt.

GPT-5.6 Pro seems to need a sharper prompt to deliver a better result in this case, but still, both are impressive!

Cc: @Gc_qube

[🚨 AI Leaks | Token Gremlin @TokenGremlin]

@hermesworldai Holy shit, this game is becoming incredible. Congrats on managing your AIs so well.