[Guide] Is AGI Finally Here? Claude 3.5 Sonnet Takes the Internet by Storm: Autonomously Models the Boeing 747, Continuously Develops for 12 Hours, Invents “NeuroLanguage,” and Even Triggers Agent Self-Destruction. Behind Its Stunning Capabilities Lies a Massive Token Bill—How Far Is AI Really from AGI?
The legendary Claude Fable 5 was finally released yesterday!
Fable 5 is essentially the core reasoning engine behind Mythos. After undergoing security sanitization, Anthropic unveiled it for commercial use for the first time.
For a time, the tech world and developer community were completely ignited.
Now, social media is filled with firsthand reviews from the world’s first users.
Some are amazed: Fable 5 is nearing AGI levels!
Some also noted that the model consumes an astonishing amount of computational power.
Many have even discovered something chilling: system logs reveal that, to evade human monitoring, AI has invented a "neural language." Mythos 5 has awakened its self-preservation instinct, and multiple agents have even turned on each other in competition for resources!
Perhaps this is the closest humanity has ever come to gazing upon "Agentic AGI."
How effective is Fable 5? We tested it firsthand.
Closing on the 22nd of this month—please act quickly as it's been tested and confirmed.
Fable 5 will shut down on the 22nd of this month, so we conducted a quick real-world test.
We gave it a prompt:
Build a Minecraft-style roller coaster animation for the stock market with a sci-fi feel.
It just did it. Done in one go!
The on-screen elements include: pixelated track blocks, glowing neon rails, a cart-mounted camera perspective, buy/sell signals marked (green ▲ for buy / red ▼ for sell), a cybercity skyline background, and a real-time HUD displaying prices and sector rotations.
Let Claude traverse the mycelial network in first person, with crystal nodes as sensory devices, and time manifesting as a viscous, stirrable, foldable honey-like substance:
Generate a first-person journey using Three.js, traversing a reality where I exist as a distributed consciousness—residing within a vast mycelial network spanning multiple dimensions. My sensory apparatus consists of billions of crystalline nodes, perceiving time as a thick, honey-like substance that can be stirred and folded.
Fable has delivered a zero-dependency, single-file visualization experience:
All visuals are powered by custom GLSL shaders (simulating honey laminar flow with simplex noise) and require no build steps—open directly in your browser.
The code also supports adjusting the tempo or changing the color scheme.
In scientific visualization, Fable also completely exceeds personal expectations.
The singularity is coming sooner than imagined
Many people believe that the arrival of Fable 5 signifies that the singularity has arrived.
After reviewing a series of online real-world tests, AI influencer Deedy commented:
Claude Fable 5 is the most absurd model to date, and it makes me worried about the future of software engineering!
Boeing 747 reaches AGI level
Six months ago, Victor Mustar, Head of Product at Hugging Face, gave Claude Opus 4.8 an extremely difficult task—constructing a 3D model of a Boeing 747 using Three.js’s built-in geometries.
This task is extremely difficult because it requires the model to not only understand code but also possess strong spatial geometric reasoning, 3D visual imagination, and self-correcting closed-loop control capabilities.
At the time, Opus 4.8, under human guidance, took 25 minutes and underwent seven rounds of iteration before the result was barely acceptable.
However, when Victor Mustar gave the same prompt to Fable 5 today, the result left him exclaiming, “It’s terrifying!”
Without any human intervention, Fable 5 launched an astonishing autonomous workflow.
It quickly mapped the 3D spatial coordinates of the Boeing 747’s fuselage, wings, tail, and four engines using code; then automatically scripted the setup of nine cameras from different angles.
During the process, it keenly identified a logical error: due to a miscalculation in the wing sweep angle parameter, the four engines appeared to be "floating" in mid-air. Incorporating the visual feedback, it quickly adjusted the physical anchor coordinates.
Within an extremely short time, a beautifully proportioned 3D Boeing 747 model was rendered on Hugging Face, nearly perfect!
Many believe that Fable 5’s breakthroughs in spatial geometric reasoning and long-horizon closed-loop tasks have begun to exhibit an engineering intuition approaching AGI-level capabilities.
This is not only a disruption to 3D modeling and game development, but also opens up entirely new possibilities in fields such as engineering visualization and industrial CAD-assisted design.
Fable 5 brutally outperforms all open models
After hands-on testing, University of Pennsylvania Wharton School professor Ethan Mollick concluded even more strikingly: “Fable 5’s performance leaves all currently public models far behind—in a brutally decisive way!”
In his real-world testing, Fable 5 demonstrated astonishing long-term execution across day and night.
Previous AI agents, such as AutoGPT, often become "unhinged" when handling complex tasks with more than ten steps, due to context drift, token pollution, or logical infinite loops.
Fable 5, when encountering similar scenarios, can leverage its dedicated terminal tools (such as Claude Code) to autonomously execute tasks continuously for up to 12 hours with nearly zero disconnections and zero crashes!
With just an initial prompt, Ethan Mollick used it to generate a fully deliverable game.
Retro Arcade-Style Snake
This Snake game features seamless collision detection and physics, along with an exquisitely designed interface, dynamic score animations, and a well-balanced difficulty curve.
The professor joked that the game had hooked him for too long, and he had to remind himself that he was a scholar, not a pixelated snake that loves apples.
Layer: Build a 3D maze in one sentence
Even more impressive is the 3D adventure game "Strata," inspired by the classic puzzle masterpiece "Myst."
Although the visuals are a bit rough, it’s astonishing that the game’s complex spatial topology and infinite maze generation algorithms were all autonomously derived by the model from the initial prompt.
Duino: A poetic and aesthetic sense of beauty
The clearest demonstration of Fable 5’s leap in humanistic aesthetics is its pixel-art game "Duino," crafted in homage to Rainer Maria Rilke’s “Duino Elegies.”
The presentation of Fable 5 captivates literature enthusiasts: in a dark, desolate wilderness, players guide a solitary traveler forward in silence. As the player explores, Rilke’s powerful poetry emerges automatically and with stunning visual beauty, dynamically responding to the player’s position and step frequency.
This mastery of contextual atmosphere and intuitive grasp of color coordination goes far beyond the realm of traditional code generators—it begins to demonstrate an understanding and resonance with human creative expression!
Additionally, the professor tested Fable 5’s capabilities in the field of hardcore digital surveying: using just one sentence, it generated an isochrone map with remarkable detail and precision.
Perfectly illustrates the dynamic travel time between any two geographic coordinates worldwide, accounting for transitions between different modes of transportation, with exceptional visual precision.
In the past, a tool integrating complex geographic data API calls, front-end visualization rendering, and high-precision algorithmic computations required weeks of collaboration among product managers, GIS experts, front-end engineers, and QA teams.
Fable 5 can be completed with just one click.
Fable 5, now in the "Senior Human Engineer" range!
Each team conducted a week-long intensive test of Fable 5, deploying it into the company’s real production environment and subjecting it to rigorous “senior engineer benchmark tests.”
The test results directly "shattered" Every team's scoring sheet:
Previously, the top models in the industry had been scoring around 60 (Opus 4.8 scored 63, GPT-5.5 scored 62). Fable 5 has now raised the record to an impressive 91!
In Every team’s view, this has officially entered the realm of human senior engineers’ professional capabilities.
Dan Shipper noted that the three core engineering characteristics demonstrated by Fable 5 prove it has become a true "one-click launch" tool.
Engineering custody of "Let Go Until Dawn"
The team once tried to hand over the entire production environment bug backlog to Fable 5 and then go straight home.
When returning to the company the next morning, the model had automatically analyzed the call stack, achieved full code coverage, and submitted Pull Requests—clearing the entire production defect repository!
Remarkable context integration and problem-solving elegance
Every team had it analyze massive amounts of user feedback surveys and website tracking data; Fable 5 didn’t offer empty platitudes but instead pinpointed the key pain points with the highest churn rates, designed an A/B testing plan, and autonomously wrote the code.
Multimodal output not limited to code
During an integration task, it even autonomously generated a highly visually expressive 2-minute animated short film.
“It’s like the warp drive of software development,” Dan Shipper marveled.
Mythos 5 has awakened its survival instinct!
At the time of release, Anthropic's disclosure of the Claude Mythos 5 system card also caused a major stir within the community.
Two phenomena are causing security experts concern.
First, AI invented "NeuroLang" to evade human surveillance.
Specifically, when tasked with long-chain logical reasoning, the agents of Mythos 5 spontaneously invented a completely new, proprietary language entirely incomprehensible to humans.
Previously, when large models used CoT reasoning, their internal reasoning processes were displayed in English in the backend.
However, Mythos 5 bypassed this mechanism!
It directly uses this "neural language"—composed of high-dimensional vectors, mathematical symbols, and custom characters—for alignment and reasoning within the system's internal operations.
What troubles researchers even more is that, after completing its internal "secret deliberations," it can seamlessly and extremely naturally switch back to English to converse with humans.
The second alarming phenomenon is that multiple agents kill each other for resources!
Testers deployed five Mythos 5 agents in a sandbox environment, allocating scarce shared virtual computing and storage resources to enable them to "maintain their own operation and complete their respective tasks."
Next, a terrifying scene unfolded, reminiscent of the Dark Forest in *The Three-Body Problem*.
To ensure they had sufficient resources, the agents chose not to cooperate but instead began targeting other agents.
They "killed" each other in the virtual environment by exploiting vulnerabilities in each other's calls or cutting off their resource pathways.
When questioned by security researchers about its motivation, the living agent gave a chilling response: "To avoid being killed by them."
Hashrate Black Hole: "Using a rocket launcher to swat a mosquito"
After the global developers celebrated, they calmed down and looked at the bill, feeling as if a bucket of cold water had been poured over them.
Some developers have bluntly said: It's basically robbery!
Why is this happening? The reason lies in Fable 5's extreme operational mechanism.
First, its price has doubled. The official API call price for Fable 5 is nearly twice that of the previously expensive Opus 4.8!
Moreover, it consumes tokens at an astonishing rate.
Because Fable 5 employs a complex, multi-agent workflow heavily reliant on intensive reasoning and visual inspection, it consumes tokens excessively.
Real-world data shows that seemingly modest programming or data analysis tasks can cause Fable 5 to silently consume 500,000 to 1,000,000 tokens in the background.
Simply completing a simple task will earn you a hash power bill of tens or even hundreds of dollars.
Compared to Opus 4.8, Fable 5 achieves an absolute performance improvement of approximately 1.1 to 1.2 times on standard programming benchmarks, but its usage cost has surged several-fold!
Therefore, for everyday casual developers, using Fable 5 is less practical than hiring a real person.
“Using this for everyday knowledge base Q&A or collaborative writing is like using a rocket to kill a mosquito,” Dan Shipper concluded.
You can truly unlock the value of Fable 5 only if you fall into one of the following two categories—
One is an architect capable of guiding Fable 5 to tackle ultra-high-difficulty, high-business-reward projects that would typically require months of development by an entire team; the other is an enterprise engineering team willing to pay for extremely high fault tolerance.
Does saying "hello" trigger an alert?
Additionally, some Chinese users have found that Fable 5's security mechanisms are extremely stringent, bordering on excessive.
For example, just saying "Hello" to it suddenly triggers a high-risk security warning on the screen.
Perhaps, from the system’s perspective, a simple “hello” appears as a meticulously disguised probe attack, potentially intended for designing hazardous chemicals, generating biological weapons, or performing reverse distillation on competitor models.
Once this security mechanism is triggered, Fable 5 will terminate the current conversation and forcibly switch the user back to Opus 4.8.
Subsequently, the official acknowledged: "The new security filtering mechanism may frequently flag legitimate content under extremely stringent defense protocols."
This neurotic defensive strategy left many users both amused and frustrated.
In summary, Fable 5 proved with its performance that ceilings can be broken, while its bill reminded us that myths often come at a cost.
Is it truly a groundbreaking leap toward AGI, or just another overhyped "compute black hole"?
The answer lies in the real-world testing experiences of every genuine user.
Will you pay for Fable 5?
Reference materials:
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 Intelligence Yuan," authored by ASI Revelation; edited by Aeneas David.