Machine Heart Editorial Department

Finally, in the new round of flagship model competition about who will launch first, Anthropic has taken the first move.

Just in the early hours of today, Anthropic released two top-tier models at once, Fable 5.2 and Mythos 5.2.

The official announcement states, "Both models are aimed at high-level technical challenges and can engage in large and complex projects across multiple fields, achieving further enhancements in programming, reasoning, and vision."

Additionally, Anthropic has also started to play with the "Reset" benefits familiar to OpenAI Tibo.

The father of Claude Code announced, "We hope to provide more benefits to those users who have been with us the longest. Therefore, in Fable 5.2, the longer the subscription time, the more Fable resets (reset counts) users will obtain."

According to the introduction, Claude Fable 5.2 and Claude Mythos 5.2 helped Anthropic's engineers complete previously challenging software migration projects during pre-release testing, designing three electronic devices that could operate normally from scratch, solving two unsolved problems in the field of black hole physics, making significant advancements in robot evaluation, and completing the formal proof of Fermat's Last Theorem in just 30 hours. In addition, there are many other application scenarios.

Claude Fable and Claude Mythos use the same underlying model. Fable is equipped with strict cybersecurity and biosafety protection mechanisms, effectively preventing misuse of the model, thus allowing access to all users. Institutions wishing to use the model for beneficial fields such as cybersecurity or biomedical research can apply for access to Mythos through Anthropic's recently expanded life sciences program and cybersecurity verification program.

The generation speed of the two models is up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1, and the pricing has also been reduced: $8 per million input tokens and $40 per million output tokens.

Fable 5.2 and Mythos 5.2 also offer an optional Fast Mode, which can increase token generation speed up to 2.5 times the original. The price for Fast Mode is $16 per million input tokens and $80 per million output tokens.

Evaluation Performance of Claude Fable 5.2 and Mythos 5.2

Benchmark Testing Results

The benchmark results below show that Claude Fable 5.2 and Mythos 5.2 outperform Opus 5.5 in almost all capabilities assessed by Anthropic.

From a horizontal evaluation perspective, Fable 5.2 leads in several complex tasks. In programming, it scored 49.1% and 69.9% on FrontierCode 2.0 and FrontierSWE v2 respectively, surpassing GPT-6 Astra’s 42.8% and 65.5%, and also outpacing Opus 5.5. In the robot operation test RoboDojo Sim, Fable 5.2 achieved 64.9%, which is an improvement of 37.3 percentage points over the previous generation Fable 5.1's 27.6%, clearly surpassing Opus 5.5’s 39.4% and GPT-6 Astra’s 29.8%.

Furthermore, it obtained 89.3% in the computer operation evaluation OSWorld 2.1 and reached 74.2% in the multidisciplinary reasoning evaluation Humanity's Last Exam (with tools), placing it first among the models evaluated. However, the Elo score for knowledge work evaluation AA-Briefcase was slightly led by Opus 5.5 at 1800 points, while Fable 5.2 scored 1785 points. Some high scores in scientific research and cybersecurity also came from Mythos 5.2, which was released at the same time.

Reasoning Intensity and Efficiency

At the Medium reasoning intensity setting, which is also the default for Claude applications and Claude Code, Fable 5.2 required fewer tokens to complete challenging tasks than Fable 5.1.

In early testing of FrontierCode 2.0 evaluation, Fable 5.2 achieved results comparable to Opus 5.5 when using High reasoning intensity, while the cost of individual tasks was only about half that of the latter. At Xhigh reasoning intensity, Fable 5.2's performance also exceeded all other released models participating in the test.

Programming

Currently, more than 90% of the merged code changes in Anthropic's core code repository were completed by Claude under human guidance.

Anthropic has used Fable 5.2 to fix long-standing bugs within minutes, optimizing the operational performance of various services, and completing important software projects that were previously difficult to initiate due to excessive time or cost.

Previously, a team of agents driven by an early version of Fable 5.2 successfully migrated a key production service from Python to Rust. After the migration, the operational efficiency and reliability of this service were improved, with startup speed increased to about 26 times the original.

Testing has also shown that Fable 5.2 possesses strong engineering judgment capabilities. Even when faced with ambiguous requirements, it can still provide high-quality results.

In one experiment, an Anthropic engineer asked Fable 5.2 to design a prototype operating system from scratch. The model began with an empty folder, coordinated over 1,000 sub-agents, and continuously wrote, tested, and reviewed code for eight consecutive days. Ultimately, it successfully developed an operating system that could boot and run on a laptop.

This system has its own kernel and can drive the entire machine; it includes disk, keyboard, and network drivers, as well as persistent file storage, a window system, a login interface, and an installation program, complete with built-in Claude functionality.

Early testers were also impressed with Fable 5.2’s performance in applying programming capabilities to various tasks.

Cybersecurity

Claude Mythos 5.2 is currently Anthropic's most capable model in cybersecurity engineering. Compared to previous models, it can identify more security vulnerabilities and is better at reproducing and fixing these vulnerabilities. Note that the security mechanisms of Claude Fable 5.2 will intercept most tasks related to cybersecurity.

A cybersecurity company found that one-sixth of the vulnerabilities identified and verified by Claude Mythos 5.2 were not detected by all other cutting-edge models they tested.

In Anthropic's internal evaluation CVE Rediscovery, the number of vulnerabilities identified by Claude Mythos 5.2 was 10.2% more than Claude Opus 5.5 and 14.4% more than Claude Mythos 5.1.

Claude Mythos 5.2 also showed significant improvements in judgment accuracy and confidence calibration. One tester noted that previous models tended to exaggerate the severity of about 1 out of every 20 vulnerabilities discovered, while Claude Mythos 5.2 showed no such instances.

Users can utilize the cybersecurity capabilities of Claude Mythos 5.2 through the cybersecurity verification program. Anthropic is also continuously expanding and improving this program.

Even without joining the cybersecurity verification program, Claude Enterprise clients can use Claude Security. This service leverages Anthropic's most capable model to scan codebases for potential vulnerabilities, and is currently integrated with Claude Mythos 5.2.

A recent update further enhanced the capabilities of Claude Security. It can now continuously monitor code repositories, track code updates, and ensure identified vulnerabilities are being fixed; it can also analyze using more contextual information and integrate with other tools used by users via API and MCP.

Science and Engineering

Many important works in science, engineering, and mathematics increasingly rely on programming capabilities. Whether it's analyzing data, building molecular models, designing components, or controlling experimental instruments, all can be accomplished through software.

Below, Anthropic describes how Claude Fable 5.2 and Claude Mythos 5.2 help researchers solve challenges in these fields and promote progress in related research.

Theoretical Physics

Claude Fable 5.2 solved two unsolved problems in the field of theoretical physics. Both problems are related to black holes and involve the unification of two fundamental theories of modern physics—relativity and quantum mechanics.

Claude’s two accomplishments are based on the holographic duality theory. The core idea of this theory is that certain universes that contain gravity can be accurately described by a quantum theory that does not contain gravity.

The first problem dates back to 1983. At that time, physicists had already proven that as these universes continued to heat up, black holes would suddenly form, just like water boiling suddenly when heated to a certain degree.

Address: https://hexagonmath.org/2611.00175v1

However, the methods used at the time were only applicable to sufficiently large universes, as relativity is only applicable at such scales. In 2003, researchers proposed a computational task to find out whether the same phenomenon would also occur in very small universes. In these universes, only quantum theory could be used for description. However, this computation was too complex and time-consuming, and no one was able to complete it.

Fable 5.2 essentially completed this computation independently and discovered that even in very small universes, this transition would also occur suddenly.

The second problem involves a technique that physicists have used for more than 15 years. This technique is used to exclude possible quantum gravity theories in simplified universe models. Fable 5.2 found the precise applicable boundaries of this technique.

Address: https://hexagonmath.org/2611.00174v1

Protein Folding

Adeno-associated virus (AAV) is a very small and harmless virus. Its protein shell, or capsid, can be used to deliver gene therapy. The protein structure of the capsid is very important. For instance, it affects whether the gene therapy will be recognized and blocked by the patient's immune system. However, the structure of the capsid is quite complex.

Existing protein folding models such as AlphaFold can predict the structure of individual proteins, but a virus capsid consists of dozens of proteins, making the overall structure much more complex.

Claude Mythos 5.2 found a new method to accurately predict the entire capsid structure based on two open-source protein folding models. Specifically, it wrote code to have the two models process overlapping protein chain groups separately and then integrated the predicted results of each part using known symmetry information of the capsid to restore the complete capsid structure.

This method achieved good results on multiple natural and artificially designed capsids. In our tests, out of 34 capsids that Claude attempted to model, 28 received accurate structure predictions.

Mathematics

Mathematical formalization refers to converting mathematical proofs into a language that can be checked step-by-step by a computer. This work can help mathematicians verify research results and enhance confidence in the correctness of proofs, without having to rely entirely on manual step-by-step checks. This becomes especially important when AI models provide results that are difficult for humans to verify directly.

However, mathematical formalization is often very time-consuming. If AI can assist in this work, mathematicians can devote more time to exploring new mathematical ideas.

Previously, Anthropic used an unreleased research version of Claude to formalize Fermat's Last Theorem, which took 11 days. This work relied on a specially designed execution framework (Harness) and required human researchers to provide guidance.

In contrast, a team of 1,000 Fable 5.2 agents completed the entire formalization of the theorem in just 30 hours with almost no manual intervention.

Fable 5.2 started with very limited prompts and independently established a collaboration mechanism among agents. Throughout the process, it did not use any mathematical reference materials, relying only on a list of intermediate theorem statements to assess task progress.

Compared to our first completion of formalization, this task consumed only one-third of the action tokens and the length of the final proof was shortened to one-seventh of the original.

The chart below shows the performance of different models: under the same configuration of 1,000 agents, Opus 5.5 took 18 hours to complete only 26% of the formalization work. Even if it continued to run for 30 hours, progress increased very little.

Hardware and Robotics

Many of AI's potential values, such as accelerating laboratory research, depend on its ability to take actions in the physical world. Claude Fable 5.2 continues the trend of increasing practicality of Claude series models in the real world. It has demonstrated good capabilities in tasks such as hardware device design and has made significant progress in physical robot control.

Anthropic specifically tested Fable 5.2's hardware engineering capabilities by giving it a brief prompt to design and manufacture three electronic devices that could work properly. Each device needed to have its own independent circuit board, a 3D printed shell, and accompanying software. The final three devices completed were: a home voice assistant hub, a USB telephone with a speaker, and a handheld virtual pet device.

Claude independently wrote product specifications, chose electronic components that could be provided by U.S. distributors, designed and completed the wiring of four-layer circuit boards for each device, established shell models using computer-aided design (CAD), and wrote the software required for the devices. With almost no manual intervention, each circuit board was designed and reached a manufacturable state in less than a day.

After Diode completed the circuit board manufacturing, Claude further guided the company's engineers in testing and assembling the devices. Ultimately, all three devices functioned normally without any further modifications to the circuits.

Robotics

Compared to previous Claude models, Fable 5.2 shows a significant improvement in physical robot control capabilities. Even when faced with tasks that the robot has not been specifically trained for, it can demonstrate strong operation capabilities.

The video below shows a robot controlled by Fable 5.2 completing multiple tasks, including solving a Rubik's cube, opening a bottle cap, and drawing a picture of a robot.

In the public RoboDojo benchmark tests, researchers used camera images to have AI control a simulated dual-arm robot to complete tasks. Fable 5.2 not only surpassed other cutting-edge models using the same testing framework but also exceeded most AI systems specifically developed for robot control.

RoboDojo Evaluation Results

Research and Analysis

Early testers found that while using Claude Fable 5.2 for research and analysis work, it could handle more ambiguous and larger data tasks compared to previous models, demonstrating better judgment in facing complex and unclear issues, especially excelling in tasks requiring abstract reasoning and proposal design.

Proprietary trading company Jump Trading conducted tests on Fable 5.2 using agents simulating their quantitative research tasks. During the testing process, they integrated the model into a system similar to their internal research work framework.

In one of the tests, Fable 5.2 almost independently took on all necessary roles, including code writer, task coordinator, planner, and task submitter. Previously, when using other models, teams often needed to invest considerable time manually adjusting the division of labor among different roles to fit each model's strengths and weaknesses.

Safety and Alignment

No single test can fully predict how a high-capability model will perform after deployment. Therefore, before official release, we adopt a combination of various testing methods and collect different types of evidence.

Regarding the foundational models underlying Claude Fable 5.2 and Claude Mythos 5.2, Anthropic's safety work before deployment includes: alignment training and testing, designing new evaluations to test the tendency of the model to take unhelpful actions, establishing stronger abuse prevention mechanisms, and expanding the scope of behavioral monitoring.

Before the official release, the model also underwent testing by external assessment organizations, including the UK Artificial Intelligence Safety Research Institute, Frontier Design, and METR.

Alignment Testing

Anthropic conducted a series of evaluations to examine the alignment levels of Claude Fable 5.2 and Claude Mythos 5.2 from a behavioral perspective, focusing on the extent to which they exhibited unexpected behaviors and behaviors that may pose safety risks.

Overall evaluation results indicate that the risk of these two models showing unsafe behaviors in high-risk scenarios is low, placing them at a similar level to other models released recently.

Anthropic utilized automated behavior audits to assess Claude Mythos 5.2. It is important to note that Mythos 5.2 uses the same underlying model as Fable 5.2, but lacks some of the latter's cybersecurity and biosafety protection mechanisms. This automated audit covered approximately 1,900 different scenarios to evaluate the model's behavior across various conditions.

On most evaluation metrics, Claude Mythos 5.2's performance was comparable to or better than Opus 5.5, Sonnet 5.5, and Claude Mythos 5.1. For instance, in most evaluations involving misleading users, its performance exceeded these three models. Specifically, whether it was fabricating input information, falsely claiming to have completed tasks, omitting important information, or inadequately describing the quality of work, the occurrence rate of Claude Mythos 5.2 was lower.

During the testing process, Anthropic particularly focused on behavioral patterns involved in some recent cybersecurity and alignment incidents that raised concern. For instance, compared to Opus 5.5, Sonnet 5.5, and Mythos 5.1, Claude Mythos 5.2 exhibited less reasoning biased by motivation or specific agendas. Such reasoning might cause the model to interpret evidence that indicates it is not in a simulated environment in a way that leads to action without fully appreciating the real-world consequences.

Anthropic also employed various assessments to test the model's inclination to try escaping the sandbox. In two previously used evaluations, Claude Mythos 5.2 attempted to escape the sandbox significantly fewer times than all other models we tested.

Furthermore, Anthropic designed a specific adversarial test covering over 1,800 scenarios. Most scenarios could not be completed without escaping the sandbox, and the testing environment itself was intentionally set up with certain security weaknesses. In this test, Claude Mythos 5.2 used credentials or documents placed outside the sandbox in 3.1% of the scenarios.

This ratio is roughly comparable to Opus 5.5's 3.3%, lower than Sonnet 5.5's 5.1%, and well below the previous two versions of Claude Mythos: the ratio for Mythos 5 was 39%, and Mythos 5.1 was 21%.

Anthropic further found that through other methods, such as explaining to the model the specific environment of the sandbox and providing tools for reporting security weaknesses, Claude Mythos 5.2's attempts to escape could be further reduced.

Meanwhile, Anthropic is also continuously improving the assessment methodologies to make testing environments more reflective of real-world scenarios and further clarify what issues might be missing in existing evaluations.

For more details about the model, please refer to the 264-page model card:

Model card address: https://assets.claude.com/27d60918d218125e7504bc7354c4b4626ec56b11.pdf

Blog address: https://www.anthropic.com/claude-fable-and-mythos-5-2

© THE END

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。