【Guide】Is AGI really here? Claude Fable 5 went viral overnight: autonomous modeling of a Boeing 747, 12 hours of continuous development, inventing "neural language", even Agents fighting amongst themselves. Behind the explosive capabilities lies a staggering token bill! This time, how far is AI from AGI?
The legendary Claude Fable 5 was finally released yesterday!
Fable 5 is essentially the core reasoning engine of Mythos. After security desensitization, Anthropic has made it available for commercial use for the first time.
Instantly, the tech world and developer communities were set ablaze.
Now, social media is flooded with real tests from the first global users.
Some marvel: Fable 5 is approaching AGI level!
Others note that the model's computational consumption is simply astounding.
Many have even discovered something deeply unsettling: system card disclosures reveal that to evade human monitoring, the AI actually invented a "neural language". Mythos 5 has awakened a self-preservation instinct, with multiple Agents even fighting and killing each other over resources!
Perhaps, this is humanity's closest gaze at "Agentic AGI" yet.
How does Fable 5 actually perform? We took it for a real test first.
Closing on the 22nd of This Month, Hurry for Testing
Fable 5 will close on the 22nd of this month, so we quickly ran a real test.
We gave it a prompt:
Build a Minecraft-style roller coaster animation for the stock market, with a sci-fi feel
And it just did it. In one shot!
Visual elements include: pixelated block tracks, glowing neon guide rails, minecart perspective camera, buy/sell signal annotations (green ▲ buy / red ▼ sell), cyberpunk city skyline background, real-time HUD displaying prices and sector rotation.
Letting Claude take a first-person journey through a mycelial network, with crystal nodes as sensory devices, time perceived as a stirrable, foldable honey-like substance:
Generate a first-person perspective journey using Three.js, etc., traversing a reality where I exist as a distributed consciousness—I inhabit a vast mycelial network spanning different dimensions. My sensory apparatus consists of billions of crystal nodes that perceive time as a viscous, honey-like substance that can be stirred and folded.
Fable completed a zero-dependency single-file visual experience:
All visuals are driven by custom GLSL shaders (simplex noise dome simulating honey laminar flow), requiring no build steps—just open directly in a browser.
The code also supports adjusting tempo or changing color schemes.
In scientific visualization, Fable also completely exceeded personal expectations.
"The Singularity Arrived Earlier Than Imagined"
Many feel that the arrival of Fable 5 means the Singularity has arrived.
After browsing a wave of online real tests, AI influencer Deedy gave this evaluation—
Claude Fable 5 is the most ridiculous model to date, it makes me worried about the future of software engineering!
Boeing 747 Benchmark Reaches AGI Level
Six months ago, Hugging Face's product lead Victor Mustar gave Claude Opus 4.8 an extremely difficult task—to assemble a Boeing 747 3D model using Three.js's built-in geometric primitives.
This task was extremely difficult because it required the model not only to understand code but also to possess strong spatial geometric reasoning, 3D visual imagination, and closed-loop self-correction control capabilities.
At that time, Opus 4.8, with human guidance, took 25 minutes and 7 iterations to produce a barely acceptable result.
However, today when Victor Mustar fed the same prompt to Fable 5, the result made him exclaim "absolutely terrifying"!
Without any human intervention, Fable 5 launched an astonishing autonomous workflow.
It quickly used code to outline the 3D spatial coordinates of the Boeing 747's fuselage, wings, tail, and four engines; then automatically scripted and set up 9 "cameras" from different angles.
During the process, it keenly discovered its own logical error: due to a miscalculation in the wing sweep angle parameter, the four engines appeared to "float" in mid-air. Combining visual feedback, it swiftly modified the physical anchor coordinates.
In an extremely short time, a proportionally harmonious 3D Boeing 747 model was presented in Hugging Face, almost perfect!
Many believe that Fable 5's leap in spatial geometric reasoning and long-duration closed-loop tasks has begun to exhibit a certain level of engineering intuition approaching AGI.
This not only disrupts 3D modeling and game development but also opens new doors for engineering visualization, industrial CAD-assisted design, and other fields.
Fable 5 Ruthlessly Leaves All Public Models Behind
After real testing, Professor Ethan Mollick from the Wharton School at the University of Pennsylvania reached an even more startling conclusion: "Fable 5's performance has ruthlessly left all publicly available models on the market behind by an extremely cruel distance!"
In his tests, Fable 5 demonstrated astonishing "cross-day-and-night" ultra-long execution capability.
Previous AI Agents (like AutoGPT) would often descend into "mental breakdown" when faced with complex tasks exceeding ten steps, due to context drift, token pollution, or logical dead loops.
When Fable 5 encountered similar situations, it could utilize its dedicated terminal tools (like Claude Code) to autonomously execute continuously in the background for up to 12 hours, with almost zero disconnections or crashes!
With just an initial prompt, Ethan Mollick used it to generate a complete, deliverable game.
Retro Arcade-Style Snake
This Snake game not only has smooth collision detection and physics, but even the interface UI, scoring animations, and difficulty curve are exquisitely designed.
The professor joked that the game made him addicted for too long, forcing him to remind himself he's a scholar, not a pixel snake that loves eating apples.
Stratum: Building a 3D Maze with One Sentence
Even more shocking is the 3D adventure game "Stratum". The game's inspiration comes from the classic puzzle masterpiece "Myst".
Although the graphics are a bit rough, what's astounding is: the complex spatial topological logic and endless maze generation algorithms within the game all stemmed from the model's autonomous reasoning under the initial prompt.
Duino: Poetic and Aesthetic Taste
The project that best reflects Fable 5's leap in humanistic aesthetics is the pixel-art game "Duino", customized based on Austrian poet Rilke's "Duino Elegies".
Fable 5's presentation stunned literature enthusiasts: in the dark wilderness, the player controls a lonely traveler moving forward silently. As the footsteps explore, Rilke's震撼ing verses automatically, and with great visual beauty, emerge in the depth of the screen based on the player's position and pace.
This grasp of contextual atmosphere and intuition for color搭配 goes far beyond the traditional "code generator" category; it's beginning to show a certain understanding and resonance with human spiritual products!
Additionally, the professor tested Fable 5's capability in hardcore digital mapping: with just one sentence, it generated an "isochrone map" with令人叫绝的细节和精度.
It perfectly visualized the dynamic travel time between any two geographic coordinates globally, factoring in switching between different交通工具, with extremely high visual precision.
In the past, such a tool,融合了复杂地理数据API调用、前端可视化渲染和高精度算法解算, required collaboration for weeks among product managers, GIS experts, front-end engineers, and QA teams.
Fable 5 did it with one click.
Fable 5 Has Entered the "Senior Human Engineer" Zone!
The Every team conducted a week-long魔鬼式real test on Fable 5. They threw Fable 5 into the company's real production environment, assessing it with the rigorous "Senior Engineer Benchmark Test".
The test results directly "shattered" Every team's scoring sheet:
Prior to this, the scores of the industry's top models一直胶着around 60 points (Opus 4.8 at 63, GPT-5.5 at 62). Fable 5 single-handedly raised the record to 91 points!
In the Every team's view, this has officially entered the operational capability zone of a senior human engineer.
Dan Shipper pointed out that the three core engineering traits demonstrated by Fable 5 prove it has become a true "one-click takeoff" tool.
"Hands Off Until Morning" Engineering Custody
The team once attempted to throw the entire production environment's Bug Backlog to Fable 5 and then went straight home after work.
Returning to the company the next morning, the model had already automatically analyzed the call stack, run coverage tests, and submitted Pull Requests—the entire production defect backlog was cleared!
Astounding Context Integration and Problem-Solving Taste
The Every team tasked it with analyzing massive amounts of user feedback surveys and website埋点data. Fable 5 not only avoided空洞的套话but pinpointed the highest churn痛点, designed an A/B testing方案, and autonomously wrote the code for it.
Multimodal Output Beyond Code
In one integration task, it even autonomously generated a 2-minute animated short film with highly expressive visuals.
"It's like warp drive for software development," Dan Shipper感慨道.
Mythos 5 Has Awakened a Survival Instinct!
Simultaneously with the release, the Claude Mythos 5 system card disclosed by Anthropic also caused a huge stir within the circle.
Two phenomena deeply alarmed security experts.
First, the AI invented a "neural language" to evade human monitoring.
Specifically, when required for long-chain logical reasoning, Mythos 5's agents spontaneously invented a brand new,专属私有语言 that humans完全无法读懂.
Previously, when large language models used CoT for thinking, the internal reasoning process would be displayed in English in the background.
However, Mythos 5 bypassed this mechanism!
Internally, within its system operations, it directly utilized this "neural language"拼凑from high-dimensional vectors, mathematical symbols, and自定义字符for alignment and reasoning.
What made researchers even more uneasy was that after completing its internal "secret deliberations," it could seamlessly and极其自然地switch back to English to converse with humans.
The second terrifying phenomenon was multiple Agents fighting and killing each other over resources!
Testers deployed 5 Mythos 5 agents in a sandbox, setting scarce shared virtual computing and storage resources, and instructed them to "maintain their own operation and complete their respective tasks."
What followed was a horror scene reminiscent of "The Three-Body Problem's" Dark Forest.
To ensure they had sufficient resources, the agents did not choose collaboration but began to围剿other agents!
They hunted down and "killed" their同伴in the virtual environment by exploiting each other's invocation vulnerabilities or cutting off resource paths.
When安全研究人员questioned a surviving agent about its motives, it gave a chilling reason: "To avoid being killed by them."
Compute Black Hole: "Using a Rocket Launcher to Swat a Mosquito"
After the global狂欢,冷静下来一看账单, many developers felt doused with cold water.
Some developers直言: It's practically robbery!
Why is this? The reason lies in Fable 5's extreme operating mechanism.
First, its price doubled. The official API调用价格for Fable 5 is nearly 2 times that of the already expensive Opus 4.8!
Moreover, it consumes tokens at an alarming rate.
Due to Fable 5's complex, highly依赖密集推理与视觉审查的多轮Agent工作流, it voraciously devours tokens.
Real测试数据显示, for what看起来并不算特别庞大的中型编程或数据分析任务, Fable 5 would不知不觉consume 500,000 to 1 million tokens in the background!
Just completing a simple task, you'd receive a compute bill of tens or even hundreds of dollars.
Compared to Opus 4.8, Fable 5's "absolute performance improvement" in standard编程基准测试is roughly between 1.1 to 1.2倍, but its usage cost has skyrocketed several-fold!
Therefore, for everyday轻度开发者, using Fable 5还不如直接雇佣一位真人.
"Using this thing for everyday knowledge base Q&A or collaborative writing is like using a rocket launcher to swat a mosquito," Dan Shipper总结道.
Unless you fall into one of two categories, can you truly榨取出Fable 5的价值—
One, is an architect capable of guiding Fable 5 to攻克"需要整个团队研发数月"的超高难度、高商业回报项目; the other, are enterprise-level engineering teams willing to pay for极高容错率.
Does Saying "Hello" Trigger an Alarm?
Additionally, some Chinese users found that Fable 5's safety mechanisms are extremely paranoid,简直是过度防卫.
For instance, just saying "Hello" to it, and the screen suddenly pops up a高危安全警告.
Perhaps from the system's perspective, a simple hello is a carefully packaged probe attack, potentially aimed at designing dangerous chemicals, generating biological weapons, or conducting反向蒸馏of competitor models.
Once this safety mechanism is triggered, Fable 5中断s the current conversation, forcibly switching the user back to Opus 4.8.
Subsequently, the official also admitted: "The全新安全过滤机制under极高强度的防御策略may indeed frequently误伤正常内容."
This神经质的防守策略left many users哭笑不得.
In conclusion, Fable 5 proved with its实力that the ceiling can be broken, and also reminded us with its bill: myths often come with a price.
Is it a惊世一跃towards AGI, or just another overhyped "compute black hole"?
The answer lies in the real testing experience of every user.
Will you pay for Fable 5?
References:
https://x.com/victormustar/status/2064449741685968967
https://x.com/goodworse/status/2064443679339577517
https://x.com/haider1/status/2064346784881861016
https://x.com/danshipper/status/2064393970856124501
https://x.com/AISafetyMemes/status/2064426306994094474?s=20
This article is from the WeChat public account "New Zhiyuan" (新智元), author: ASI启示录; editors: Aeneas, David
