product
2026-11-10 06:54
1,000 AI Agents completed the formal proof of Fermat's Last Theorem in 30 hours, worked continuously for 8 days, and built an operating system from scratch that can run on a laptop.
Just now, Anthropic officially released two new models, Claude Fable 5.2 and Mythos 5.2. In addition to the above achievements, the new model also solves two black hole physics problems, increases the generation speed by up to 30%, and reduces the API price by 20%.
Interestingly, just two months after Anthropic CEO Dario Amodei publicly called for slowing down AI development, Anthropic has once again launched a stronger cutting-edge model. Anthropic, meanwhile, is preparing for an IPO that could value the company at $1 trillion to $1.5 trillion.
While calling for slowing down AI, while sprinting for a trillion-dollar IPO, what new thing did Anthropic come up with this time?
1,000 Agents worked for 8 days to build an operating system from scratch
According to Anthropic's positioning, Claude Fable 5.2 and Mythos 5.2 are flagship models specifically designed for complex technical challenges, focusing on improving programming, reasoning, visual understanding, and scientific research capabilities.
Both use the same underlying model, with the main difference being security restrictions. Fable 5.2 adds anti-abuse mechanisms for network security and biological research and is open to ordinary users; Mythos 5.2 retains more complete professional capabilities and is mainly provided to institutions that have passed the review.
Judging from the official benchmark tests, the performance of the new model is quite impressive.
In the FrontierCode 2.0 test, which measures AI Agent programming prowess, Fable 5.2 achieved a score of 49.1%, surpassing Claude Opus 5.5's 45.5%, GPT-6 Astra's 42.8%, and Gemini 4 Carbon's 41.1%.
In the FrontierSWE v2 software engineering test, Fable 5.2 scored 69.9%, an increase of 13.6 percentage points compared to the previous generation. In the FrontierCode 2.0 test, Fable 5.2 achieved results close to the highest inference intensity of Opus 5.5 with lower inference intensity, and the single-task cost was about half of the latter.
The latest list from the third-party evaluation agency Artificial Analysis also confirms this.
The Claude Fable 5.2 tops the Intelligence Index v5 with 65 points, 5 points ahead of the Gemini 4 Carbon, 6 points ahead of the Claude Opus 5.5, and takes first place in 5 out of 9 tests. The main advantages focus on scientific reasoning, knowledge understanding and programming abilities. However, the average single-task cost in the strongest inference mode reaches $26.5, which is approximately 2.6 times that of Opus 5.5.
However, what deserves more attention than the running scores are the actual engineering cases displayed by Anthropic.
An engineer approached Fable 5.2 with a request to design an operating system from scratch. The model started from an empty folder, autonomously coordinated more than 1,000 sub-agents, and worked continuously for 8 days to complete code writing, testing and review.
Eventually, Claude succeeded in developing an operating system that could boot and run on a laptop.
According to the official display, the system has its own kernel, supports disk, keyboard and network drivers, and also has file storage, window interface, login screen, installer and built-in Claude function.
Although it is still an experimental prototype system, AI has been able to autonomously coordinate thousands of agents to complete such complex software engineering projects, which is enough to reflect the changes in its capabilities.
In another practical case, an early version of Fable 5.2 also migrated a key Anthropic production service from Python to Rust, making it about 26 times faster to start.
After its release, AI entrepreneur Matt Shumer commented that Fable 5.2 is the strongest model he has ever used. It can even complete complex projects that other models have been unable to achieve in the past. However, it is sometimes "overly responsible" and even blocks normal work tasks.
Netizen Kevin Ngo also posted a 3D game produced by Fable 5.2 that can be played directly in the browser. Players can travel through 8 dream worlds by boat;
Another developer, Flavio Adamo, also showed off the 3D works generated by the model, saying that its three-dimensional creation capabilities are amazing.
In addition, netizen Synopsis also used the classic "Pelican Riding a Bike" test to challenge Fable 5.2, showed off the work generated by the model, and directly called it a new generation of BOSS.
Anthropic also revealed that more than 90% of the merged code changes in the company's core code base are currently written by Claude under human guidance.
From fixing simple bugs to autonomously planning and executing large-scale engineering tasks that last for days, AI programming is gradually gaining the ability to independently undertake complex projects.
Complete the formal proof of Fermat's last theorem in 30 hours, and also solve the black hole physics problem
If the improvement in programming capabilities is still expected, then the scientific research results announced by Anthropic are even more surprising.
Fermat's Last Theorem is one of the most famous difficult problems in the history of mathematics. Although mathematician Andrew Wiles completed the proof as early as the 1990s, converting it into a formal proof that computers can verify step by step is still an extremely complex task.
Anthropic previously used an unreleased Claude research model, which took 11 days to complete under the guidance of human researchers.
Today, a team of 1,000 Fable 5.2 Agents completed the formal proof of Fermat’s Last Theorem in just 30 hours, with almost no manual intervention.
The model establishes its own collaboration mechanism in the experiment, and the length of the generated proof is shortened to one-seventh of the previous one, and the operation token consumption is also reduced to one-third of the original one. As a comparison, Claude Opus 5.5 with the same configuration ran for 30 hours and only completed about 27% of the proof work.
In addition to mathematics, Fable 5.2 also addresses two open problems in theoretical physics related to black holes.
One involves the phase transition mechanism when a black hole forms. Scientists discovered as early as 1983 that black holes in certain theoretical models can suddenly form as temperature changes, but whether this conclusion can be extended to smaller quantum scales has been lacking complete calculation results.
Claude basically completed the calculations that had been unfinished for a long time independently and came to the same sudden conclusion. In another study, models determined the precise boundaries of an analytical approach to quantum gravity theory.
At the same time, Mythos 5.2 also developed a new method to predict the structure of complete viral capsids by combining two existing open source protein structure prediction models, and successfully accurately modeled 28 of the 34 capsids tested.
Although relevant scientific research results still need further verification, AI has begun to participate in more professional research tasks, from mathematical proofs to black hole physics to protein structure prediction.
Design three electronic devices in one go, and the robot begins to understand instructions
Claude's abilities also began to extend to real-world hardware design.
In an experiment Anthropic collaborated with circuit board design and manufacturing company Diode, researchers used just a prompt word to get Fable 5.2 to design three working electronic devices: a home voice assistant hub, a USB conference speaker, and a handheld electronic pet.
Throughout the process, Claude independently wrote product specifications, selected electronic components, designed a four-layer PCB circuit board, built a 3D printed housing model, and developed supporting software.
According to Anthropic, the circuit board designs for each device reached a manufacturable state in less than a day. After Diode completed manufacturing, Claude guided engineers through testing and assembly. In the end, all three devices worked successfully without modifying the electrical design.
In terms of robot control, the progress of Fable 5.2 is equally obvious.
In the RoboDojo Sim dual-arm robot simulation test, Fable 5.2 achieved a success rate of 64.9%, exceeding the 27.6% of the previous generation Fable 5.1 and the 29.8% of GPT-6 Astra.
In the official demonstration, the physical robot controlled by Claude also completed tasks such as unscrewing bottle caps, restoring Rubik's cubes, and drawing robot patterns.
Coincidentally, netizen Tekon Burns also took out a real Rubik's Cube test Fable 5.2, uploading only 6 photos, the model gave a set of solutions with very few steps, and did not follow the common human problem-solving idea of first spelling the white cross and then restoring the white surface.
From text and code on the screen, to real-life circuit boards and robots, Claude is trying to translate his reasoning abilities into actual physical operations.
Network security continues to upgrade, but AI R&D capabilities are actively restricted
Compared to Fable 5.2, the focus of Claude Mythos 5.2 is still on network security.
Anthropic first launched Mythos in April this year. Considering its powerful vulnerability discovery and exploitation capabilities, it was only open to selected institutions at that time, and then launched the Fable series with increased security restrictions for public use.
With this upgrade, Mythos 5.2 was able to rediscover 89.6% of known software vulnerabilities in CVE Rediscovery internal testing, up from 78.3% in the previous generation.
One security firm also found that about one-sixth of the vulnerabilities confirmed to be discovered by Mythos 5.2 were not discovered by other leading-edge models it tested.
As its capabilities improve, Anthropic has further strengthened its model security measures. In a sandbox escape test that included more than 1,800 scenarios, the proportion of Mythos 5.2 using credentials or files outside of the sandbox dropped to 3.1%, down from 21% in the previous generation.
But what deserves more attention is Anthropic’s proactive restrictions on new models.
Officials made it clear that during the training process, the new model was specifically limited in improving its capabilities in areas such as AI accelerator Kernel optimization and large language model architecture improvements.
Therefore, while Fable 5.2 and Mythos 5.2 have significantly increased overall capabilities, their performance remains roughly at the level of Claude Opus 5.5 on tasks that may directly accelerate the development of next-generation AI.
Anthropic emphasized that the relevant adjustments did not affect capabilities such as ordinary Agent programming, hardware development, and traditional machine learning research.
In addition, Anthropic has also strengthened its measures to prevent the illegal distillation of model capabilities, claiming that related technology can reduce the amount of data obtained by malicious distillers by 90% to 99%.
Price reduced by 20%, Claude Pro has hidden benefits
Claude Fable 5.2 has been officially launched today, covering the Claude platform and cloud services such as AWS, Google Cloud and Microsoft Azure. Developers can call the new model through claude-fable-5-2.
Compared with the previous generation, the standard API price of the new model is reduced by about 20%, charging $8 per million input tokens and $40 per million output tokens.
Both models also launch an optional Fast Mode, which can generate tokens up to 2.5 times faster than the standard mode. The price is accordingly increased to US$16 per million input tokens and US$80 per million output tokens.
Anthropic said that the Fable series is mainly aimed at complex engineering, mathematics and scientific research tasks, and it is still recommended to use Claude Opus 5.5 first for daily programming and knowledge work.
To promote the new model, Anthropic is also launching a limited-time trial event.
From November 9th to 16th local time, qualified Claude Pro users and Team standard seat users can use the Fable series models within 50% of the original weekly usage quota without purchasing additional usage quota, and will receive a free quota reset opportunity.
Max, Ultra and some Team and Enterprise premium seat users can get 1 to 3 usage quota resets based on the paid subscription length.
It should be noted that the reset function will only restore the amount that has been consumed and will not permanently increase the usage limit of the subscription plan.
As for Mythos 5.2, it is still primarily available to verified cybersecurity agencies and life science research organizations.
Calling for slowing down AI while sprinting towards trillion-dollar IPO
The timing of Anthropic’s new model release is quite subtle.
In September of this year, Anthropic CEO Dario Amodei publicly called for slowing down the speed of AI development, hoping that the entire industry would be more cautious in handling the rapidly improving model capabilities and their potential risks.
Now, in just the past two months, Anthropic has unveiled a new, more capable, cutting-edge model.
In response, Dianne Penn, head of research product management at Anthropic, responded in an interview with CNBC, "What we call is to control the pace of development, not to suspend development."
Anthropic stated that the new model underwent extensive safety and alignment testing before its official release, and was evaluated by external agencies such as the British AI Safety Institute and METR.
However, investment in security is also facing increasingly realistic business pressures.
According to CNBC, Anthropic is making final preparations for its IPO, and management is about to launch a listing road show with a target valuation of US$1 trillion to US$1.5 trillion, lower than the US$2 trillion target previously reported by the market.
According to Anthropic's updated prospectus, in the nine months ended September 30 this year, Anthropic achieved revenue of $33.24 billion, a year-on-year increase of approximately 1,200%. For a company that is about to sprint towards a trillion-dollar IPO, technological leadership means greater business opportunities, but also means increasingly unbearable costs of deceleration.
Perhaps the most subtle paradox of the AI race is here. Everyone knows that they should apply the brakes, but they are all afraid of being the first to slow down.
Shan Zong
Hello, World Hello, Word
A total of 508 articles have been published
This article comes from the column
Next article2026-11-10 08:45
Tesla Cybertruck, which once claimed to have over one million orders, has sold only 63,000 vehicles in total, and has been in mass production for less than four years, officially announced the suspension of production
Previous article2026-11-09 22:56
Maserati will jointly build cars with Zunjie. The first model may be a two-door GT, which will be officially put into mass production next year.
Aifaner App
Love style, make the future within reach
Digital trendy brands focusing on tomorrow’s products
Advanced tools, prophets first, AIGC’s inspiration guide
A new era of car manufacturing, tomorrow’s travelers
Explore new lifestyles in the city and be your guide to tomorrow’s life
Copyright © Guangzhou Liwo Zhiyuan Investment Management Partnership (Limited Partnership) 2008 - 2026. Use of Aifaner website content for commercial purposes requires permission. Non-commercial use is licensed under CC BY-NC 4.0.
All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.
