AGI fehlt nur noch der letzte Schritt

One should still remember that in April, Anthropic released a model called Mythos.

Just from the name, one can tell how impressive it is: Mythos.

At that time, it is said to have helped 50 corporate customers find more than ten thousand highly dangerous security vulnerabilities, which shocked the entire industry.

This news temporarily led to a massive crash in network security stocks. One should still keep this in mind.

Due to its excessive power and the fear of misuse – “too dangerous for the public” – it was not made accessible to the general public.

Only last night did Anthropic add a security classifier to the Mythos model and officially launch Fable 5 online.

The uncensored Mythos 5 is currently only accessible to about 200 strictly vetted institutions such as the White House, network security protection teams, and the Butterfly Project.

Such caution can't help but make one think of the currently popular AI animation “Angel Machine.”

Is there an “angel” in the cage?

Even if it isn't the case now, it will be soon.

01

Based on the test data officially released by Anthropic and the first practical reports from corporate partners, the power of Fable 5 can be described with the word “stunning.”

Let's first look at the test values.

In the automatic programming evaluation ranking SWE - Bench Pro, Claude Fable 5 has a pass rate of 80.3%. Its “mother” Opus 4.8 has 69.2%, GPT - 5.5 has 58.6%, and Gemini 3.1 Pro only has 54.2%.

In the advanced code evaluation, Fable 5 reaches 29.3%, Opus 4.8 reaches 13.4%, and GPT - 5.5 only reaches 5.7%.

...

The difference is so great that it's as if someone suddenly took out a machine pistol in the era of cold weapons.

In all other tests such as software development, independent scientific hypotheses, drug molecule design, model distillation and boundary compression, long - term context understanding, etc., Fable 5 is almost always in the first place.

Those who want more detailed information can watch videos about it.

Now let's look at the practice.

The payment giant Stripe used Fable 5 in an early test. They had a legacy codebase of 50 million lines that needed to be comprehensively migrated. According to estimates, such a restructuring would take at least two months even for a professional team.

After the problem was handed over to Fable 5, it independently created a plan, checked its progress, and corrected errors by itself. Within one day, the migration of the 50 - million - line codebase was completed.

Such performance can't simply be described with the word “powerful.”

In a narrow sense, Fable 5 has already achieved artificial general intelligence (AGI) at the level of the digital economy.

The reason is that it shows a real “long - term agent ability.”

Whether it's GPT - 5.5 or Gemini 3.5, not to mention other less powerful models, they are essentially just “answer providers.”

One has to prompt it to make it do something.

If it gets into a dead end, it can only throw an exception and say: “Sorry, I'm just a language model.”

Although it is called a tool, the user still has to think deeply and guide the AI system step by step to get the desired result. This is not an easy task.

Fable 5 with its internal goal - oriented logic is different.

As in Stripe's test, when the user gives it a difficult long - term task that runs in three steps:

Creating a subtask tree;

Planning and executing various tools (web search, database query, Python sandbox environment);

Self - reflection: If something doesn't work, another way is immediately chosen.

Except for setting the task and accepting the result, the user doesn't need to intervene further.

As a productivity tool, it is already very perfect.

But it is still not the real AGI.

The power of Fable 5 is based on the fact that the underlying codebase, scientific literature, etc. still have a mathematical logic and structural definition.

The reason why it doesn't get lost in long - term tasks is that it has overcome the problem of “attention decay in long texts.” When processing complex tasks with millions of tokens, it can always stay focused on the core goal.

But if you throw it into a completely chaotic physical reality that lacks digital rules and is not fully understood by humans, it will have logical gaps due to the lack of a “foundation.”

If we refer to the “Five - Level Standards for Artificial Intelligence” proposed by OpenAI (Level 1: Chatbot; Level 2: Thinker; Level 3: Agent; Level 4: Innovator; Level 5: Organization).

Opus 4.8 is on the way from Level 2 to Level 3, while Fable 5 is firmly at Level 3 and is advancing towards Level 4.

It took 43 days from Opus 4.7 to 4.8, and only 11 days from 4.8 to Fable 5.

How long will it take to reach Level 4? Given Anthropic's increasingly rapid update rate, it will probably be possible this year.

Even the final Level 5, according to optimistic estimates, will only take 18 - 24 months. It's really just one step away.

This speed is too fast. This is also the main reason why security restrictions need to be introduced.

02

In the 《System Card》 and the RSP evaluation report released by Anthropic together with the model, Mythos 5 has shown extremely dangerous signals in two capabilities.

First, the Fable/Mythos base model has reached the CB - 1 level in the chemical and biological evaluation.

This means that the model has the ability to “synthesize non - new biological/chemical weapons end - to - end and guide their production” and even give suggestions for genetic sequence changes to optimize the transmission efficiency of a highly dangerous virus.

If a terrorist with a bachelor's - level biological knowledge gets the unleashed Mythos 5, he can obtain complete instructions by constantly prompting the model on how to bypass raw material control, how to set up a simple P3 laboratory in the basement, and how to synthesize highly lethal pathogens.

Second: Network attacks and exploitation of security vulnerabilities.

Even in the very early tests, Mythos 5 has shown that it is able to autonomously find and exploit core vulnerabilities in critical infrastructures (such as power plants, financial cashier systems, hospital networks). Within seconds, it can generate an attack script tailored to the zero - day problem.

When Mythos was developed in April this year, there were rumors that it had found more than ten thousand highly dangerous security vulnerabilities among 50 initial partners.

...

Under these two circumstances, it would be too dangerous to leave Mythos 5 to the general public.

One has to lock this monster in a cage.

After two months, Anthropic built a two - layer cage.

First: Silent degradation routing mechanism.

Anthropic has installed a completely independent, extremely sensitive, and high - precision classifier AI system in front of Fable 5.

If the user enters a complex query that may involve network attacks, biochemistry, or a hidden hint to query the model weights, the classifier immediately triggers an alarm and silently redirects the session in the background to the older Opus 4.8 to answer the question.

Second: Data storage.

Anthropic and Amazon jointly announced last night that all data streams that call the Mythos model, regardless of whether it is a first - or third - party platform, must be stored for 30 days.

Why?

Because real hackers or terrorists are usually very intelligent. They won't directly ask in a conversation: “How do you make a bomb?” but break the question into 100 seemingly harmless basic questions.

The 30 - day full - scale data monitoring is intended to detect such “salami - slicing” misuse scenarios through pattern recognition, which are not recognizable in a single conversation.

As Dario Amodei warned in previous public statements: “The probability that artificial intelligence poses a catastrophic threat to humanity is as high as 25%.”

In order to comply with the internal guidelines 《Responsible Scaling Policy》 (RSP) and 《Frontier Compliance Framework》 (FCF), Anthropic has to put shackles on this monster itself.

Thus, Fable 5 was born.

03

Now let's talk about the price.

The official price released by Anthropic is: $10 per million input tokens and $50 per million output tokens.

This is too expensive.

In today's corporate agent tasks, a chained logic with “multiple considerations” is often used to achieve high accuracy. One run can consume 20 million input tokens and output 5 million processed lines of code.

In total, a single task costs $450.

In addition, Anthropic has announced that the test window for the Mythos model in the existing personal subscriptions (Claude Pro) will be permanently closed on June 22, 2026.

If private users use it for work in the future, they will spend several dozen dollars within seconds.

Although the price will generally decrease with technological development, by then it will no longer be the most powerful model.

The situation is very clear: The latest and most powerful models will become luxury goods that ordinary people can't afford.

Naturally, this is understandable for Anthropic, which focuses on the B2B market.

The problem is that Google recently announced a price war.

Why does Anthropic dare to raise the price when its competitors' prices are falling?

Because the token price is only apparent, the return on investment is the essential thing.

Corporate customers don't care how much a kilowatt - hour of electricity or a token costs. As long as the AI system can complete the entire project work without errors, they are willing to pay a premium.

What's even more important is that today's network security war has completely turned into a conflict between AI systems.

Since models at the Fable/Mythos level can immediately find security vulnerabilities in systems, companies and government institutions have no choice but to pay a high price to Anthropic for private network security defense with Mythos 5 to protect themselves from attacks.

Simply put, it's like a protection fee: I've created the most feared sword (Mythos 5). I'm afraid it might hurt someone, so I sell it with a scabbard to the public (Fable 5). At the same time, I sell the unrestricted sword to the defense forces so that they can use it to intercept other sword developments.

Protection against AI threats will become an indispensable expense item for every large company.

This will lead to the high budgets in the B2B market being more concentrated on Anthropic, while cheap models that are only suitable for creating office correspondence and emails will have to compete for customers in the C2C market with low profit margins.

It is foreseeable that the global network security sector will...