Everyone should remember that in April, Anthropic released a model named Mythos.
Just from the name, you can tell how powerful it is—mythical.
At the time, it was reported that over 10,000 critical vulnerabilities were identified for 50 enterprise clients, shaking the entire industry.
This news once caused a sharp decline across the entire cybersecurity stock sector—many of you likely still remember it.
Due to its overwhelming power, concerns about potential misuse, and being "too dangerous to release," it is not made available to the public.
Until last night, Anthropic added a safety classifier to the Mythos model and officially launched Fable 5.
As for the uncut Mythos 5, access is currently limited to approximately 200 organizations that have undergone rigorous review, such as the White House, Cyber Defenseers, and the Transparent Wings Project.
Such caution makes it hard not to think of the recently popular AI-animated series "Angel Engine."
Is the one locked in the cage the "angel"?
Even if it’s not now, it’s not far off.
01
According to official test data released by Anthropic and real-world reports from its first enterprise partners, Fable 5’s capabilities are nothing short of astonishing.
First, check the benchmark score.
On the automated programming evaluation leaderboard SWE-Bench Pro, Claude Fable 5 achieves an 80.3% pass rate, while its predecessor Opus 4.8 scores 69.2%; GPT-5.5 reaches 58.6%, and Gemini 3.1 Pro stands at just 54.2%.
Leading code evaluation: Fable 5 scores 29.3%, Opus 4.8 scores 13.4%; GPT-5.5 scores just 5.7%.
……
The gap is like someone suddenly pulling out a machine gun in the era of cold weapons.
In all other areas—software engineering, independent scientific hypotheses, drug molecule design, model distillation and extreme compression, long-context understanding, and more—Fable 5 ranks first in nearly every test.
For specific details, you can watch the video.
Now let’s look at the practical application.
Payment giant Stripe conducted an early test with Fable 5. They had a legacy codebase of up to 50 million lines that required a full-codebase migration. According to assessments, such a large-scale refactoring would take at least two months even if assigned to a professional team.
As a result, after feeding the task to Fable 5, it planned the entire process on its own, monitored its progress, and corrected errors independently. In just one day, it completed the migration of 50 million lines of code.
This level of performance cannot be adequately described by just the two words "powerful."
From a narrow perspective, Fable 5 has already achieved AGI at the level of the digital economy.
The reason is that it demonstrates genuine "long-range agency."
Whether it's GPT-5.5 or Gemini 3.5, let alone other smaller large models, they are essentially all "responding."
You kick it once, and it takes one step.
Hit a dead end—it can only throw an exception and plead, "Sorry, I'm just a language model."
It may seem like a tool, but users still need to think deeply and guide the AI step by step to achieve the desired results—it’s not easy.
Fable 5, with its internalized goal-oriented logic, is different.
Like Strip's test, when users give it a complex long-term task, break it down into three steps:
Build a subtree of tasks;
Schedule different tools (web search, database retrieval, Python sandbox environment);
Reflect on yourself, realize it’s not working, and immediately switch to another path.
People no longer need to stand by and micromanage—just assign tasks and receive results.
As a productivity tool, this is already perfect.
But it is still entirely different from true AGI.
Fable 5's strength is built upon the underlying mathematical logic and structural definitions that still exist in its codebase, scientific literature, and other foundational resources.
It remains focused during long-term tasks by overcoming the challenge of "long-text attention decay," maintaining alignment with core objectives even when processing complex tasks involving millions of tokens.
But once it is thrown into the chaotic, digitally unregulated physical reality of human society—where even humans themselves don’t fully understand the rules—it still suffers logical fractures due to a missing foundation.
If measured against OpenAI's "Five Levels of Artificial Intelligence" (Level 1: Chatbot; Level 2: Reasoner; Level 3: Agent; Level 4: Innovator; Level 5: Organization).
Opus 4.8 is advancing from Level 2 to Level 3, while Fable 5 has firmly established itself at Level 3 and is exploring Level 4.
From Opus 4.7 to 4.8, it took 43 days; from 4.8 to Fable 5, it took only 11 days.
How long will it take to reach Level 4? Given Anthropic’s increasingly rapid update pace, it’s likely to be achieved within this year.
Even the final Level 5, with optimistic estimates, requires only 18 to 24 months—truly just one step away.
This speed is too fast, which is precisely why security restrictions are necessary.
02
In Anthropic’s accompanying System Card and RSP evaluation report for the model, Mythos 5 exhibited extremely concerning signals in two capabilities.
First, the Fable/Mythos base model has achieved CB-1 level in chemical and biological evaluations.
This means the model has end-to-end capabilities to synthesize and guide the production of non-novel biological or chemical weapons, even providing recommendations for genetic sequence modifications to optimize the transmission efficiency of high-risk viruses.
If a terrorist with a basic undergraduate background in biology obtained unrestricted access to Mythos 5, they could, through persistent prompting, obtain complete guidance on evading raw material regulations, setting up a makeshift P3 laboratory in a basement, and synthesizing highly lethal pathogens.
Second, network attacks and exploitation of vulnerabilities.
During early-stage testing, Mythos 5 demonstrated the ability to autonomously identify and exploit critical infrastructure vulnerabilities—such as those in power plants, financial clearing systems, and hospital networks—generating targeted zero-day exploit scripts within seconds.
When Mythos was first developed in April this year, reports revealed that over 10,000 critical vulnerabilities were identified for its initial 50 partners.
……
In both of these scenarios, directly releasing Mythos 5 to the public is far too risky.
This beast must be locked in a cage.
Two months later, Anthropic's cage has two levels.
First, the silent downgrade routing mechanism.
Anthropic has deployed a completely independent, highly responsive, high-precision classifier AI on the frontend of Fable 5.
When a user inputs a complex prompt that may involve cyber defense, biochemistry, or attempts to extract model weights, the classifier immediately triggers an alert and silently routes the conversation in the background to the older Opus 4.8 model for response.
Second, data retention.
Anthropic and Amazon jointly announced last night: All traffic calling the Mythos model, whether on first-party or third-party platforms, must enforce a mandatory 30-day data retention policy.
Why?
True hackers or terrorists typically have high IQs and won't directly ask, "How do I make a bomb?" in a conversation; instead, they break the question down into a hundred seemingly harmless basic questions.
Thirty days of comprehensive data monitoring is designed to identify "salami-slicing" malicious abuse through pattern recognition—abuse that cannot be detected in single conversations.
As Dario Amodei previously warned in public: "There is a full 25% probability that AI could lead to catastrophic risks for humanity."
To comply with Anthropic’s internal Responsible Scaling Policy (RSP) and Frontline Compliance Framework (FCF), Anthropic must personally put the chains on this behemoth.
Thus, Fable 5 was created.
03
Let’s talk about price again.
Anthropic's official pricing is $10 per million input tokens and $50 per million output tokens.
Too expensive.
Current enterprise-level agent tasks, in pursuit of high accuracy, often employ a chain-of-thought logic involving "multiple rounds of reasoning," with a single pass consuming up to 20 million input tokens and outputting 5 million lines of revised code.
That comes out to $450 per task.
Additionally, Anthropic has announced that the access window for the Mythos model included in existing individual subscriptions (Claude Pro) will be permanently closed on June 22, 2026.
In the future, if individual users actually use it for work, dozens of dollars will be spent in the blink of an eye.
Although, with technological advancements, its price will eventually come down, by then it will no longer be the strongest.
The current situation is clear: the most advanced large models have become luxuries that ordinary people simply cannot afford.
Of course, this is entirely reasonable for Anthropic, which focuses on the B2B market.
The problem is that not long ago, Google loudly announced it was engaging in a price war.
Why does Anthropic dare to raise prices when competitors are cutting prices to grab market share?
Because token prices are artificial, the return rate is what truly matters.
Enterprise clients don’t care about the cost per kilowatt-hour or the price of a single token—as long as AI can flawlessly execute the entire engineering workflow, they’re eager to pay the premium.
More importantly, today’s cybersecurity battles have become outright battles between AI and AI.
Since Fable/Mythos-level models can instantly identify system vulnerabilities, enterprises and government institutions have no choice but to pay a premium to Anthropic for a private, on-premises defense service for Mythos 5 to prevent attacks.
In simple terms, it’s protection money: I created the most terrifying sword (Mythos 5), and out of fear of causing harm, I sold the sheathed version to the public (Fable 5), while simultaneously selling the unrestricted sword to defense agencies so they can use it to intercept swords others are developing.
Defending against AI threats will become a mandatory expense for every large enterprise.
This will directly cause enterprise market budgets to concentrate further on Anthropic, while low-cost models suitable only for drafting documents and sending emails will be forced to compete fiercely in the low-margin consumer market.
It is foreseeable that the global cybersecurity sector will soon undergo an AI-driven valuation reassessment.
At the same time, "one-person businesses" will soon become an increasingly common phenomenon.
04
Features built-in task budget allocation, supports memory work and context management, capable of remembering, revising, and restarting like a human, and can independently handle the entire lifecycle from requirement documentation to code delivery.
The emergence of Fable 5 and Mythos 5 is less a technological update to large models and more a rite of passage marking the full maturation of AI industry specialization.
The AI market has initially bid farewell to the idyllic era of "free trials for everyone."
The most advanced computing power and deepest intelligence will be prioritized as strategic productive resources and directed toward infrastructure, research, and B-end application domains that generate the greatest commercial value.
This is a celebration of explosive productivity and a winter for the labor market.
This article is from the WeChat official account "Gelong," authored by Wan Lianshan.