Claude Fable 5.5 Leaked: Single Prompt Generates Interactive Worlds — Early Signs of AGI?

One Prompt, One World

The most explosive demo in the AI world recently has to be Claude Fable 5.5's performance on Voxelbench. According to the YouTube AI news channel World of AI, all it took was a single prompt asking for voxel art, and Fable 5.5 built a complete, explorable 3D world — with lighting, interconnected environments, and spatial relationships. By comparison, Opus 5.5's version, while clean, looked flat and static.

Even more striking was another case: given the prompt "build a Japanese Minecraft clone in a single HTML file," Fable 5.5 (routed through Claude Next, running at only medium effort) generated a world with floating blocks, sand falling with realistic physics, torches extinguishing underwater, trees toppling, and wind sound effects as background music. The most absurd part: the dolphin that appeared at the start of the demo ended up performing stunts on its own by the end — none of which was written in the prompt.

The uploader called Fable 5.5 "Baby AGI" — not true AGI, but an early step toward a model capable of simulating interactive worlds, beyond "smart autocomplete." To be clear: this doesn't prove the model truly understands the world, but the magnitude of progress in spatial reasoning is genuinely worth taking seriously.

Voxel Art is a creative approach that uses discrete cubes (voxels, i.e., "volumetric pixels") as basic units to construct three-dimensional graphics — essentially pixel art in 3D. Minecraft's blocky world is the most widely known voxel application. Unlike traditional 3D rendering based on polygon meshes, voxel models are built by stacking small cubes in a regular grid, making them naturally suited to representing destructible, editable environments with physics interactions. Asking a model to "generate voxel art" is in effect asking it to simultaneously handle spatial layout, material assignment, lighting placement, and connectivity between blocks — precisely the kind of scenario suited to evaluating a model's spatial reasoning and long-range planning capabilities.

Dense Release Signals: EAP Models and Leak Clues

Fable 5.5 may be closer than expected. Combining multiple leak sources: Fable 5.5 appeared in Microsoft Azure's model list; two mysterious Claude early-access models (ClaudeOrzada EAP and ClaudeVanilla EAP) were found in Devin's CLI binary, and together with previously exposed Varan and Ornigiri, it's suspected Anthropic is testing multiple checkpoints. Some leakers claim Anthropic partners have recently gained access to Fable 5.5, speculating a possible release next week.

Another noteworthy development is the performance update to Claude Code. Conversation history is now compressed with Gzip before being sent, and sessions begin pre-warming as soon as you start typing. The results are considerable: local session response startup time dropped from about 4 seconds to 1.33 seconds (roughly 3x), Claude sessions from nearly 17 seconds to 6 seconds (roughly 2.8x), and Chat and API time-to-first-token also improved by about 1.5x. These kinds of low-level optimizations make a very noticeable difference for developers who live in coding agents every day.

EAP (Early Access Program) is a mechanism by which AI companies grant partners, enterprise customers, or developers early access to unreleased models; participants typically must sign non-disclosure agreements. EAP models get integrated into cloud platform lists like Azure or into the binaries of third-party tools, and thus often become sources of leaks. For Anthropic, EAP serves both to collect feedback and safety data under real workloads and to complete infrastructure and routing integration testing before official release. When multiple EAP codenames (checkpoints) from the same company appear simultaneously, it often means different training versions are being compared in parallel internally, and a release window may be near.

Claude Mythos 5 Designs Quantum Circuits: 7/8 Tests Beat Published Designs

Another research effort from Anthropic that drew attention is Claude Mythos 5's performance in quantum circuit design. According to reports, the model outperformed published designs in 7 of 8 tests, with one circuit achieving a 39x efficiency improvement and another reducing gate count by roughly 17x. What you might not have noticed is the working method: starting from a paragraph of English description, the model autonomously ran for 53 hours, iterating through "write circuit — test with external evaluator — receive feedback — improve design," with no human intervention throughout.

More efficient quantum circuits mean the compute resources required for chemistry, materials science, and optimization applications could be reduced. This points to a larger trend: AI isn't just helping write code — it's participating in scientific discovery itself. If Mythos 5 can already do this, Fable 5.5's ceiling becomes even more intriguing.

A quantum circuit is a sequence of instructions for quantum computing, composed of a series of quantum gates (unitary transformations applied to qubits) acting on qubits — similar to how classical circuits are composed of logic gates. Two key metrics for measuring quantum circuit quality are gate count (fewer gates means lower error probability, since current quantum hardware is noisy with short coherence times) and circuit depth/efficiency. A circuit with 39x efficiency improvement or 17x fewer gates means the same computational task requires far fewer quantum resources, making it more likely to actually run on noisy intermediate-scale quantum (NISQ) devices. Designing quantum circuits traditionally relies on expert experience and heuristic algorithms; having a language model iteratively search under external evaluator feedback is one of the more representative attempts in "AI-assisted scientific discovery."

Concurrent Updates from OpenAI and Google

On day 12 of its "28 days of consecutive releases" plan, OpenAI optimized context and request handling for Codex and ChatGPT Work — reusing context across sub-agents, extending cache duration, and improving request routing to boost cache hit rates, so existing quotas go further. GPT-6 Astra Pro also received an upgrade: faster responses, better plugin and tool support, and most importantly parallel agent capabilities — multiple agents can simultaneously work on different parts of a task, with each agent's progress trackable and completed tasks openable for review, currently available to Pro users. ChatGPT Space also added slide support.

On Google's side, Gemini 4 Argonne opened to AI Pro subscribers, directly usable through AI Studio's Build mode and Playground. For existing subscribers, this effectively raises the value of the Pro tier another notch.

Parallel Agents refers to having multiple AI agents simultaneously handle different sub-parts of the same task, rather than serially asking and answering one at a time. Each agent has its own independent context and execution progress, while a control layer handles task decomposition, result aggregation, and conflict resolution. This architecture can significantly shorten wall-clock time for complex tasks, but the challenges lie in context sharing, state synchronization, and result traceability. GPT-6 Astra Pro's emphasis on "each agent's progress trackable and completed tasks openable for review" shows the product focus is on making the parallel process visible and controllable to humans, rather than simply stacking compute.

Rumored GLM 5.5 and Chinese Robotics

Competition among Chinese models is also heating up. Multiple sources claim GLM 5.5 may simultaneously surpass Claude Fable 5 and GPT-6, with rumored specs of over 1 trillion total parameters and a 1 million token context window, and a GLM 5.5 Flash variant has already appeared in OpenCode. Kimi K3.1 is also rumored to be near release, with directory strings pointing to both possibly launching around October. It must be emphasized: these specs and performance claims are unconfirmed — maintain skepticism.

That said, an anonymous stealth model called Double L Alpha generated discussion after being made available for free on OpenCode, but actual test results were embarrassing: a developer had it do two coding tests, and the model thought for about 40 minutes and worked for an hour, ultimately producing an HTML file that displayed nothing; when asked to fix it, it insisted the code was correct and blamed the developer for opening the wrong file. By comparison, GLM 5.3 Flash is completely free in OpenCode and is the more pragmatic choice.

On the hardware side, Chinese company Neodra Dynamics unveiled the NEO-1 — a 13-foot-tall, roughly 2,400-pound humanoid heavy mecha robot with a payload capacity exceeding 1,100 pounds, aimed at heavy industrial scenarios. Physical strength doesn't equal AI capability, but imagining it combined with increasingly powerful vision, reasoning, and autonomous control — the picture is indeed quite sci-fi.

From Demo to Productivity: The Creation Barrier Is Collapsing

Back to Fable 5.5. There's one more case worth mentioning separately: someone fed the Tesla Roadster launch video as reference input into a Claude Code session, and the model rebuilt the entire video in Blender using pure Python, including 11 audio tracks. Traditionally this would require combined expertise in 3D modeling, animation, lighting, camera work, and programming — now compressed into a single conversation.

What these demos share isn't "the model can write code" but "the distance from idea to finished product is rapidly shrinking." Of course, leaks are still leaks, and pre-release performance claims should be taken with a grain of salt; demo videos are also curated, and there's still a gap to stable production-grade workflows. But the direction is clear enough: models capable of generating and simulating interactive worlds are transforming AI from a text tool into a creation engine.