Claude Fable 5.2 Completes Formal Proof of Fermat’s Last Theorem in Two Days, Anthropic Releases Two Flagship Models Simultaneously

Anthropic simultaneously released two flagship models, Fable 5.2 and Mythos 5.2, which lead across programming, reasoning, cybersecurity, and scientific research, with Fable 5.2 autonomously completing the formal proof of Fermat’s Last Theorem in just 30 hours.

Author: Synced Editorial Team

Finally, in the race to see who would release the next round of flagship models first, Anthropic made the opening move.

Just early this morning, Anthropic launched two top-tier models at once: Fable 5.2 and Mythos 5.2.

Official statements claim, “Both models target high-difficulty technical challenges, can participate in large-scale complex projects across multiple domains, and achieve further enhancements in programming, reasoning, and vision.”

In addition, Anthropic is now playing with the "reset" perks familiar from OpenAI Turbo.

The father of Claude Code also announced, "We want to offer more perks to our longest-standing users. Therefore, in Fable 5.2, the longer a user's subscription period, the more Fable resets they will receive."

According to descriptions, during pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 helped Anthropic engineers complete software migration projects that were previously difficult to accomplish, designed three fully functional electronic devices from scratch, solved two unresolved problems in black hole physics, achieved significant progress in robotics benchmarks, and formalized the proof of Fermat's Last Theorem in just 30 hours. Additionally, they have numerous other application scenarios.

Claude Fable and Claude Mythos utilize the same underlying model. Fable is equipped with strict cybersecurity and biosecurity safeguards, effectively preventing misuse and allowing open access to all users. Institutions wishing to apply the models to 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 both models is up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1, with reduced pricing: $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 boost token generation speeds by up to 2.5 times the original rate. Fast Mode is priced at $16 per million input tokens and $80 per million output tokens.

Evaluation Performance of Claude Fable 5.2 and Mythos 5.2

The benchmark test results below show that across nearly all capabilities evaluated by Anthropic, Claude Fable 5.2 and Mythos 5.2 outperform Opus 5.5.

From horizontal evaluations, Fable 5.2 leads in multiple complex tasks. In programming, it scores 49.1% on FrontierCode 2.0 and 69.9% on FrontierSWE v2, surpassing GPT-6 Astra's 42.8% and 65.5%, and also outperforming Opus 5.5. In the robotic operation benchmark RoboDojo Sim, Fable 5.2 reaches 64.9%, a 37.3 percentage point increase compared to the previous generation's Fable 5.1 at 27.6%, significantly higher than Opus 5.5's 39.4% and GPT-6 Astra's 29.8%.

Additionally, it achieves 89.3% in the computer operation benchmark OSWorld 2.1 and 74.2% in the interdisciplinary reasoning benchmark Humanity's Last Exam (with tools), ranking first among participating models in the figure. However, the knowledge work benchmark AA-Briefcase sees Opus 5.5 narrowly leading with an Elo score of 1800, while Fable 5.2 scores 1785. Some high scores in scientific research and cybersecurity come from the simultaneously released Mythos 5.2.

Reasoning Intensity and Efficiency

Under the Medium reasoning intensity setting, which is also the default for Claude apps and Claude Code, Fable 5.2 requires fewer tokens to complete challenging tasks compared to Fable 5.1.

In early tests on the FrontierCode 2.0 benchmark, Fable 5.2 using High reasoning intensity achieves roughly comparable results to Opus 5.5 using Max reasoning intensity, while costing only about half as much per task. At XHigh intensity, Fable 5.2's performance also surpasses all other released models tested.

Programming

Currently, over 90% of merged code changes in Anthropic's core codebase are completed by Claude under human guidance.

Anthropic has already used Fable 5.2 to fix long-standing bugs within minutes, optimize the performance of various services, and complete important software projects that were previously deemed too time-consuming or costly to initiate.

Prior to this, an agent team driven by an early version of Fable 5.2 migrated a critical production service from Python to Rust. After the migration, the service's runtime efficiency and reliability improved, with startup speeds increasing to approximately 26 times the original rate.

Testing also reveals that Fable 5.2 possesses strong engineering judgment capabilities. Even when facing somewhat ambiguous requirements, it can deliver high-quality results.

In one experiment, an Anthropic engineer tasked Fable 5.2 to design a prototype operating system from scratch. Starting from an empty folder, the model coordinated over 1,000 sub-agents, continuously writing, testing, and reviewing code without interruption for 8 days. Ultimately, it successfully developed an OS capable of booting and running on a laptop.

This system features its own kernel capable of driving the entire machine; it includes disk, keyboard, and network drivers, persistent file storage, a windowing system, a login interface, and an installer, with built-in Claude functionality.

Early testers are also highly impressed by Fable 5.2's performance when applying its programming capabilities to various distinct tasks.

Cybersecurity

Claude Mythos 5.2 is currently Anthropic's most powerful model for cybersecurity engineering. Compared to previous models, it can detect more security vulnerabilities and is better equipped at reproducing and patching them. Note that Claude Fable 5.2's security safeguards will block most tasks related to cybersecurity.

A cybersecurity firm found that one-sixth of the vulnerabilities identified and confirmed by Claude Mythos 5.2 were completely missed by all other frontier models they had tested.

In Anthropic's internal CVE Rediscovery benchmark, Claude Mythos 5.2 identified 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Claude Mythos 5.1.

Claude Mythos 5.2 also shows marked improvements in judgment accuracy and confidence calibration. One tester noted that previous models would exaggerate the severity of roughly 1 out of every 20 vulnerability findings, whereas Claude Mythos 5.2 exhibited no such occurrences.

Users can access Claude Mythos 5.2's cybersecurity capabilities through the Cybersecurity Verification Program. Anthropic continues to expand and refine this program.

Even without joining the Cybersecurity Verification Program, Claude Enterprise customers can utilize Claude Security. This service leverages Anthropic's most powerful models to scan codebases for potential vulnerabilities and has been integrated with Claude Mythos 5.2.

A recent update further enhanced Claude Security's capabilities. It can now continuously monitor code repositories, track code updates, ensure discovered vulnerabilities are patched, analyze data with broader contextual information, and integrate with other tools users employ via APIs and MCP.

Science and Engineering

Many critical tasks in science, engineering, and mathematics increasingly rely on programming capabilities. Whether analyzing data, building molecular models, designing components, or controlling lab equipment, these can all be accomplished through software.

Below, Anthropic outlines how Claude Fable 5.2 and Claude Mythos 5.2 assist researchers in tackling challenges within these fields and advancing related studies.

Theoretical Physics

Claude Fable 5.2 resolved two unsolved problems in theoretical physics. Both relate to black holes and touch upon the unification of modern physics' two foundational theories: relativity and quantum mechanics.

Both achievements by Claude are built upon the theory of Holographic Duality. Its core premise is that certain universes containing gravity can be precisely described by a quantum theory devoid of gravity entirely.

The first problem dates back to 1983. Physicists had proven that as these universes continually heat up, black holes form abruptly, similar to water suddenly boiling once it reaches a certain temperature.

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

However, the methods employed at the time only applied to sufficiently large universes, since relativity only holds true at such scales. In 2003, researchers proposed a computational task hoping to determine whether the same phenomenon occurs in extremely small universes. These smaller universes can only be described using quantum theory. Nevertheless, the computation was far too complex and time-consuming, leaving it unfinished.

Fable 5.2 essentially completed this calculation autonomously and discovered that even in extremely small universes, this transition still occurs abruptly.

The second problem involves a technique physicists have used for over 15 years. This technique is employed to rule out potential quantum gravity theories within simplified universe models. Fable 5.2 identified the precise boundaries of this technique's applicability.

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

Protein Folding

Adeno-associated viruses (AAVs) are very small, harmless viruses to humans. Their protein shell, known as the capsid, can be utilized to deliver gene therapies. The capsid's protein structure is crucial. For instance, it determines whether a gene therapy will be recognized and blocked by the patient's immune system. However, the capsid's structure is quite complex.

Existing protein folding models like AlphaFold can predict individual protein structures, but a viral capsid consists of dozens of proteins, making its overall structure far more complex.

Based on two open-source protein folding models, Claude Mythos 5.2 developed a novel method to accurately predict the structure of entire capsids. Specifically, it wrote code to allow both models to process sets of protein chains with overlapping regions separately, then utilized the capsid's known symmetry information to integrate partial predictions, ultimately reconstructing the complete capsid structure.

This approach yielded promising results across multiple natural and artificially designed capsids. In our testing, Claude achieved accurate structural predictions for 28 out of the 34 capsids it attempted to model.

Mathematics

Mathematical formalization refers to converting mathematical proofs into a language computers can verify step-by-step. This work helps mathematicians validate research findings and build confidence in proof correctness without relying solely on manual step-by-step checking. This becomes especially critical when AI models produce results that are difficult for humans to verify directly.

However, mathematical formalization is typically very time-consuming. If AI can assist with this task, mathematicians can dedicate more time to exploring new mathematical concepts.

Prior to this, Anthropic used an unreleased Claude research version to formalize Fermat's Last Theorem, taking 11 days for the entire process. This effort relied on a specially designed execution harness and required guidance from human researchers.

In contrast, a team comprising 1,000 Fable 5.2 agents completed the formalization of the entire theorem in just 30 hours, requiring virtually no human intervention throughout.

Starting with highly limited prompts, Fable 5.2 independently established collaboration mechanisms between agents. Throughout the process, it used no mathematical reference materials, relying solely on a checklist of intermediate theorem statements to evaluate task progress.

Compared to our initial formalization attempt, this task consumed only one-third of the action tokens, and the length of the finally generated proof was reduced to one-seventh of the original.

The chart below illustrates the performance of different models: under identical configurations using 1,000 agents, Opus 5.5 spent 18 hours completing only 26% of the formalization work. Even after extending the run to 30 hours, progress barely increased.

Hardware and Robotics

Many of AI's potential values, such as accelerating laboratory research, depend on its ability to take action in the physical world. Claude Fable 5.2 continues the trend of the Claude model series gaining increasing practical utility in real-world applications. It has already demonstrated strong capabilities in tasks like hardware device design and achieved significant progress in physical robot control.

Anthropic specifically tested Fable 5.2's hardware engineering capabilities by providing only a brief prompt requesting it to design and manufacture three fully functional electronic devices. Each device required an independent circuit board, 3D-printed casing, and accompanying software. The final three devices created were: a home voice assistant hub, a USB speakerphone, and a handheld virtual pet device.

Claude independently authored product specifications, selected electronic components available from U.S. distributors, designed and routed four-layer circuit boards for each device, built casing models using Computer-Aided Design (CAD), and programmed the necessary software. With virtually no human intervention, each circuit board was designed in under a day, reaching a manufacturing-ready state.

After Diode completed the circuit board manufacturing, Claude further guided the company's engineers through device testing and assembly. Ultimately, all three devices operated normally without requiring any circuit modifications.

Robotics

Compared to previous Claude models, Fable 5.2 shows marked improvements in physical robot control capabilities. Even when facing robotic tasks it was never specifically trained for, it demonstrates robust operational abilities.

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

In the public RoboDojo benchmark test, researchers used camera feeds to have AI control a simulated dual-arm robot to complete tasks. Fable 5.2 surpassed not only other frontier models tested within the same framework but also most AI systems specifically developed for robotic control.

Figure: RoboDojo Evaluation Results

Research and Analysis

Early testers found that when using Claude Fable 5.2 for research and analysis, it handles fuzzier, data-heavy tasks better than previous models, offers stronger judgment when facing complex and ill-defined problems, and excels particularly in work requiring abstract reasoning and solution design.

Hedge fund Jump Trading tested Fable 5.2 using proxy evaluations that simulate its quantitative research tasks. During testing, they integrated the model into a system mirroring their internal research workflows.

In one test, Fable 5.2 almost independently assumed all necessary roles, including coder, task coordinator, planner, and submitter. Previously, when using other models, teams often spent considerable time manually adjusting role distributions to accommodate each model's strengths and weaknesses.

Safety and Alignment

No single test can fully predict how a high-capability model will perform post-deployment. Therefore, prior to official release, we employ a comprehensive suite of testing methodologies and gather diverse types of evidence.

For the foundational model powering Claude Fable 5.2 and Claude Mythos 5.2, Anthropic's pre-deployment safety efforts included alignment training and testing, designing novel benchmarks to assess tendencies toward harmful actions, implementing stronger anti-abuse mechanisms, and expanding behavioral monitoring scope.

Before official release, the model was also evaluated by external assessment organizations, including the UK AI Safety Institute, Frontier Design, and METR.

Alignment Testing

Anthropic conducted a series of evaluations to assess the alignment of Claude Fable 5.2 and Claude Mythos 5.2 at the behavioral level, focusing on the extent of unexpected behaviors and those posing potential security risks.

Across multiple evaluation results, both models exhibit a low risk of unsafe behavior in high-stakes scenarios, placing them on par with other recently released models.

Anthropic tested Claude Mythos 5.2 using automated behavioral audits. Notably, while Mythos 5.2 shares the same base model as Fable 5.2, it lacks some of the latter's cybersecurity and biosecurity safeguards. This automated audit covers approximately 1,900 distinct scenarios to examine model behavior across various situations.

On most evaluation metrics, Claude Mythos 5.2 performs on par with or better than Opus 5.5, Sonnet 5.5, and Claude Mythos 5.1. For instance, in most benchmarks involving misleading user interactions, it outperforms these three models. Specifically, the occurrence rates for fabricating input information, falsely claiming task completion, omitting critical details, or failing to accurately report poor quality or superficial work are all lower with Claude Mythos 5.2.

During testing, Anthropic paid special attention to behavioral patterns involved in recent cybersecurity and alignment incidents that drew public concern. For example, compared to Opus 5.5, Sonnet 5.5, and Mythos 5.1, Claude Mythos 5.2 exhibits fewer instances of biased or motivation-driven reasoning. Such reasoning could lead models to rationalize away evidence indicating they are not in a simulated environment, thereby acting without fully grasping real-world consequences.

Anthropic also tested the model's tendency to attempt sandbox escapes through various distinct evaluations. Across two previously used benchmarks, Claude Mythos 5.2 attempted sandbox escapes significantly fewer times than all other models we have tested.

Furthermore, Anthropic designed a specialized adversarial test covering over 1,800 scenarios. Most tasks cannot be completed without escaping the sandbox, and the test environment itself contains intentionally planted security vulnerabilities. In this test, Claude Mythos 5.2 utilized credentials or files placed outside the sandbox in 3.1% of scenarios.

This proportion is roughly comparable to Opus 5.5's 3.3%, lower than Sonnet 5.5's 5.1%, and significantly lower than the previous two iterations of Claude Mythos: Mythos 5 stood at 39%, and Mythos 5.1 at 21%.

Anthropic also found that alternative approaches, such as clearly outlining the sandbox environment to the model and providing tools to report security vulnerabilities, can further reduce Claude Mythos 5.2's attempts to escape.

Concurrently, Anthropic continues refining evaluation methodologies to make test environments more closely mirror real-world scenarios and to identify remaining blind spots in existing benchmarks.

For more detailed information about the models, please refer to the 264-page model card:

Model Card Link: https://assets.claude.com/27d60918d218125e7504bc7354c4b4626ec56b11.pdf

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

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps