Token Gremlin

I believe everything we’re creating with AI should have one single purpose: helping every living being live a better life. It’s that simple. And I sincerely hope everyone else feels the same way, because the technology itself is incredibly promising, and I love it.

With all the progress I have seen in game design from GPT-5.6 Sol and Fable 5, I believe we will have a AAA game created entirely, or almost entirely, by AI within the next five years. Seriously... It is absolutely incredible.

GPT-5.6 Sol generated this entire real-time cloth simulation in code using Three.js, WebGPU, and TSL. No game engine, no downloaded models, no pre-made fabric textures, and no baked animation. The silk material, structural behavior, wind, lighting, stage, and camera sequence are all procedural and running live in the browser. This is an excellent example of how far AI-generated visual systems are moving beyond static frontend work. Cc: @AdemVessell

Fable 5 built this real-time ocean scene directly in the browser using Three.js, WebGPU and TSL, with no engine and no baked assets. The interesting part is not just that the water looks good. The whole system is being simulated: wave spectrum, GPU FFT, foam generation and refracted-ray caustics, all running live. This is exactly where frontier models start becoming genuinely useful for technical graphics work. Not by replacing years of rendering research, but by compressing the path from complex math to a working interactive result. A small, focused scene, but technically very serious. Cc: @AdemVessell

@kristalksai @posi_posi8 This one in particular was developed by a colleague, but I really appreciate the compliment. 🫡

@TokenGremlin @posi_posi8 Stumbled on your work this evening and I have to say, it's really awesome to see. Will be following with a keen eye 🤝

Fable 5 also handled the render, camera choreography, and final video generation around the scene. That is a much broader creative pipeline than simply producing a 3D environment. Cc: @posi_posi8

GPT-5.6 Sol has been working on Power Atlas for more than 16 hours straight without hitting a single rate limit. It is researching sources, reconciling conflicting data, verifying results and continuously expanding a global map of energy and digital infrastructure with almost no intervention from the user. That is the part worth paying attention to. Not the pretty map by itself, but the fact that the model can keep doing useful, structured work for this long without collapsing into loops, forgetting the objective or requiring constant babysitting. You can access the project and test it yourself here: power-atlas.sarvesh-kapre.chatgpt.site Cc: @_cyberhusky

GPT-5.6 Sol and Fable 5 just one-shotted a playable F-16 combat simulator in a browser. Sol handled the systems thinking, flight behavior, radar logic, combat architecture and how everything should connect. Fable 5 turned that reasoning into working Three.js code. One model designed the machine. The other built it. In a few hours, they had real terrain, forests, mountains, a lake, live telemetry, radar tracking, target locking, guns, missiles and a complete HUD running inside a browser tab. This is obviously not a finished game. Mission design, balancing, progression, realism and long-term retention are still deeply human problems. But the technical barrier just collapsed. A project that once required a small team and weeks of prototyping can now emerge from one person orchestrating two frontier models over an afternoon. We are no longer watching AI generate isolated pieces of code. We are watching specialized models behave like an engineering team. Cc: @explosss1ve

OpenAI now employs more than 400 former Apple employees. That talent spans engineering, industrial design, software, hardware, manufacturing, operations, and supply chain expertise. Combined with elite researchers, product leaders, and the teams brought in through io, OpenAI is no longer just an AI lab. It is assembling the human infrastructure required to design chips, operating systems, interfaces, consumer devices, and tightly integrated hardware and software experiences. A phone, a keyboard, headphones, a wearable, or an entirely new category of personal computer are all within its technical reach. OpenAI has built a war machine made of some of the most experienced people in the technology industry.

@mrflashy < Fable > GPT

@TokenGremlin Which is which

Fable 5 vs GPT-5.6 Sol in Blender MCP is becoming a very interesting comparison. Fable 5 currently looks more consistent out of the box for this kind of 3D scene generation, while GPT-5.6 Sol seems much more sensitive to how the prompt is structured. Both are impressive.

@heyamalantony My internal sources have not confirmed the size of GPT-6 yet, not even approximately. But I am expecting something in the range of 8 to 10 trillion parameters.

@TokenGremlin it won’t be the same size as Mythos. GPT 6 will be a 5-7T params model with a new pre-trained base

GPT-5.6 Sol is a fascinating model because it appears to represent the absolute limit of what OpenAI could extract from the GPT-5.x architecture through post-training. It is smaller than Fable-class models, at roughly 3–4 trillion parameters, but extremely optimized. Better tool use, stronger long-horizon execution, improved UI generation, more reliable coding, and much better agentic behavior all came from pushing the same underlying family to its practical ceiling. GPT-6 should be a very different story. Instead of squeezing more capability out of an existing base, OpenAI appears to have trained a substantially larger model from scratch, with enough raw scale to compete directly with Fable and Mythos-class systems. Sol shows how far exceptional post-training can take a model. GPT-6 should show what happens when that same post-training is applied to a much more powerful foundation.

Fable 5 + Blender MCP is seriously impressive. This is a great example of how the model can build a coherent 3D environment instead of just isolated assets. The street layout, buildings, spatial consistency, and overall scene structure are already there. Workflows like this are making AI-driven 3D scene generation look more and more practical. Cc: @posi_posi8

No, things are not tense internally at Anthropic. To be honest, they have never been better. They are impressed by GPT-5.6 Sol, but they expected more. The team remains steady, working on Fable 5.1 without much pressure to race against the clock. I sincerely hope GPT-6 shakes that structure and creates real fear, but for now, that is still just faith on my part.

@ggg78g89 Fake, bro. 🫡

🚨SCOOP: MY Friend at Anthropic says things are VERY tense internally. Dario's running tough meetings — GPT-5.6 Sol is strong and Grok 4.5 is right on Opus's heels. Pulling Fable from subs on July 12 would trigger mass cancellations (why keep Max for Opus 4.8?), so they're now pushing to keep Fable 5 in subs permanently.

GPT-5.6 Sol Ultra was given a full set of character references and asked to rebuild the character in Blender, rig it, create three animations, export a real GLB, and build a complete Three.js showcase around it. The entire pipeline was completed in 2 hours and 11 minutes. The final model still loses some of the original proportions, costume structure, and material fidelity, but going from reference images to a rigged, animated 3D character running in the browser is already a very strong result. Cc: @givros

GPT-5.6 is substantially smaller than Fable, with roughly 3 to 4 trillion parameters. OpenAI squeezed every last drop out of it through the post-training of the GPT-5.x line, and the result was exceptional. But now it is time to flip the game and train a cosmic silicon monstrosity that matches Fable in scale, or exceeds it. They know this, and they have already done it.

Actually, OpenAI bet was post-training (small/mid sized models) and Anthropic was pre-training (of the largest you can train model). OpenAI is playing catch up because post-training enhances model ability in verifiable domains (used as judges in post training) but leaves the model constrained when connecting different domains of knowledge (n-dimensional random walks during inference). However, that should be said, they KNOW how to do post training very well.

GPT-6 is finished and is reportedly dramatically better than Fable 5. OpenAI chose to train a new, larger model from scratch. Anthropic, by contrast, is continuing to train the model it already has, which means we will likely see more Fable variants, such as Fable 5.1. The two companies are taking different approaches right now because one already has a massive model in Fable/Mythos, while the other needed to build one with GPT-6.

@mktpavlenko @voxelbench Yes!

@TokenGremlin @voxelbench 145 points is a meaningful gap, not just a marginal lead.

GPT-5.6 Sol has reached the top spot on @voxelbench. With a 2302 Elo rating, the model now leads Fable 5 by 145 points. The result is especially interesting because the examples show strong range across voxel-style generation, with good structure, consistency, and visual clarity. A very strong showing for GPT-5.6 Sol.

He takes scamming to a whole new level