Claude Fable 5.2 Completes Formal Proof of Fermat’s Last Theorem in Two Days, Anthropic Releases Two Flagship Models Simultaneously
Author: Machine Learning Editorial Team
Finally, in the new round of competition over who will release the flagship model first, Anthropic has made its move.
Earlier this morning, Anthropic released two top-tier models in one go: Fable 5.2 and Mythos 5.2 have arrived.
According to official statements, "The two models are designed to tackle highly challenging technical tasks, capable of participating in large-scale complex projects across multiple domains, while achieving further improvements in programming, reasoning, and vision."
Additionally, Anthropic has started offering the "Reset" benefits familiar to OpenAI's Tier 1 users.
The creator of Claude Code also announced, "We hope to provide more benefits for our long-standing users. Therefore, in Fable 5.2, the longer a user has been subscribed, the more Fable resets (number of 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 achieve, designed three fully functional electronic devices from scratch, solved two unresolved problems in black hole physics, made significant progress in robotic evaluations, and completed a formal proof of Fermat's Last Theorem in just 30 hours. Additionally, they offer numerous other application scenarios.
Claude Fable and Claude Mythos share the same underlying model. Fable is equipped with strict cybersecurity and biosecurity protection mechanisms to effectively prevent model abuse, making it available 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.
Both models generate 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 increase token generation speed by up to 2.5 times. The price for Fast Mode is $16 per million input tokens and $80 per million output tokens.
Benchmark Performance of Claude Fable 5.2 and Mythos 5.2
The benchmark results below show that across almost all capabilities evaluated by Anthropic, Claude Fable 5.2 and Mythos 5.2 outperform Opus 5.5.
In horizontal evaluations, Fable 5.2 takes the lead 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 outperforming Opus 5.5. In the robotic manipulation benchmark RoboDojo Sim, Fable 5.2 reaches 64.9%, an improvement of 37.3 percentage points over the previous generation Fable 5.1's 27.6%, significantly higher than Opus 5.5's 39.4% and GPT-6 Astra's 29.8%.
Additionally, it achieves 89.3% on the OSWorld 2.1 computer usage benchmark and 74.2% on the multidisciplinary reasoning benchmark Humanity's Last Exam (with tools), both ranking first among the models in the chart. However, in the knowledge work benchmark AA-Briefcase, Opus 5.5 narrowly leads with an Elo score of 1800, compared to Fable 5.2's 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 the Claude app 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 performance roughly comparable to Opus 5.5 using Max reasoning intensity, while costing only about half per individual task. Under XHigh reasoning intensity, Fable 5.2's performance also surpasses all other released models included in the tests.
Programming
Currently, over 90% of the merged code changes in Anthropic's core codebase have been completed by Claude under human guidance.
Anthropic has already used Fable 5.2 to fix long-standing bugs in minutes, optimize the performance of various services, and complete some critical software projects that were previously too time-consuming or costly to initiate.
Previously, an agent team powered by an early version of Fable 5.2 migrated a key production service from Python to Rust. After the migration, the service's execution efficiency and reliability were improved, and its startup speed increased to approximately 26 times the original.
Tests also revealed that Fable 5.2 possesses strong engineering judgment capabilities. Even when facing relatively vague requirements, it can deliver high-quality results.
In one experiment, an Anthropic engineer tasked Fable 5.2 with designing a prototype operating system from scratch. Starting from an empty folder, the model coordinated over 1,000 sub-agents to continuously write, test, and review code for eight uninterrupted days. Ultimately, it successfully developed an operating system 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, along with built-in Claude capabilities.
Early testers were also impressed by Fable 5.2's performance when applying its programming capabilities to various different tasks.
Cybersecurity
Claude Mythos 5.2 is currently Anthropic's most powerful model for cybersecurity engineering. Compared to previous models, it can identify more security vulnerabilities and is better at reproducing and patching them. Note that the safety guardrails of Claude Fable 5.2 will block most cybersecurity-related tasks.
A cybersecurity company found that one-sixth of the vulnerabilities identified and confirmed by Claude Mythos 5.2 were completely missed by all other frontier models they tested.
In Anthropic's internal benchmark CVE Rediscovery, Claude Mythos 5.2 discovered 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Claude Mythos 5.1.
Claude Mythos 5.2 also shows significant improvement in judgment accuracy and confidence calibration. One tester noted that previous models would exaggerate the severity of roughly 1 in every 20 vulnerability findings, whereas Claude Mythos 5.2 has not exhibited this behavior.
Users can access the cybersecurity capabilities of Claude Mythos 5.2 through the Cybersecurity Verification Program. Anthropic continues to expand and refine this program.
Even without joining the Cybersecurity Verification Program, Claude Enterprise customers can use Claude Security. This service leverages Anthropic's most capable model to scan codebases for potential vulnerabilities and has now integrated 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 discovered vulnerabilities are patched; it can also analyze with richer 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 laboratory equipment, everything can be accomplished through software.
Below, Anthropic explains how Claude Fable 5.2 and Claude Mythos 5.2 help researchers tackle challenges in these fields and drive progress in related studies.
Theoretical Physics
Claude Fable 5.2 has resolved two unsolved problems in theoretical physics. Both relate to black holes and touch upon the unification challenge between modern physics' two foundational theories: general relativity and quantum mechanics.
Both achievements 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 entirely devoid of gravity.
The first problem dates back to 1983, when physicists proved that as these universes heat up, black holes suddenly form, much like water suddenly boiling once it reaches a certain temperature.
Link:https://hexagonmath.org/2611.00175v1
However, the methods used at the time were only applicable to sufficiently large universes, as relativity only applies at that scale. In 2003, researchers proposed a computational task hoping to determine whether the same phenomenon would occur in extremely small universes, where only quantum theory applies. Yet, the computation was too complex and time-consuming, leaving it unfinished.
Fable 5.2 essentially completed this computation independently and found that this transition occurs suddenly even in extremely small universes.
The second problem involves a technique that physicists have used for over 15 years. This technique is employed to rule out candidate quantum gravity theories within simplified universe models. Fable 5.2 identified the precise boundaries for this technique's applicability.
Link:https://hexagonmath.org/2611.00174v1
Protein Folding
Adeno-associated virus (AAV) is a very small virus harmless to humans. Its protein shell, known as the capsid, can be used to deliver gene therapies. The protein structure of the capsid is crucial; for instance, it affects whether the gene therapy will be recognized and neutralized by the patient's immune system. However, capsid structures are also quite complex.
Existing protein folding models like AlphaFold can predict the structure of individual proteins, but a viral capsid consists of dozens of proteins, making the overall structure far more complex.
Building on two open-source protein folding models, Claude Mythos 5.2 has developed a new method to accurately predict the structure of an entire capsid. Specifically, it wrote code to have the two models separately process protein chain groups with overlapping regions, then utilized the capsid's known symmetry information to integrate the partial predictions, ultimately reconstructing the complete capsid structure.
This approach yielded excellent results across several natural and artificially designed capsids. In our tests, Claude successfully generated 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 that computers can step-by-step verify. This work helps mathematicians validate research findings and build confidence in proof correctness without relying entirely on manual step-by-step checks. This becomes even more critical when AI models produce results that are difficult for humans to directly verify.
However, mathematical formalization is typically very time-consuming. If AI could assist with this work, mathematicians could dedicate more time to exploring new mathematical ideas.
Previously, Anthropic used an unreleased research version of Claude to formalize Fermat's Last Theorem, taking 11 days in total. This work 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.
Starting from a very limited prompt, Fable 5.2 autonomously established collaboration mechanisms among the agents. Throughout the process, it did not use any mathematical reference materials, relying solely on a list of intermediate theorem statements to assess task progress.
Compared to our initial formalization process, this run consumed only one-third of the action tokens, and the final generated proof length was reduced to one-seventh of the original.
The chart below illustrates the performance of different models: given the same configuration of 1,000 agents, Opus 5.5 spent 18 hours completing only 26% of the formalization work. Even when allowed to run until the 30-hour mark, progress barely increased.
Hardware and Robotics
Much of AI's potential value, such as accelerating laboratory research, depends on its ability to take action in the physical world. Claude Fable 5.2 continues the trend of the Claude series models becoming increasingly practical in real-world scenarios. It has already demonstrated solid capabilities in tasks like hardware device design and made significant progress in controlling physical robots.
Anthropic specifically tested Fable 5.2's hardware engineering capabilities by giving it only a brief prompt to design and manufacture three fully functional electronic devices. Each device required an independent circuit board, a 3D-printed enclosure, and companion software. The three completed devices were: a home voice assistant hub, a USB speakerphone, and a handheld virtual pet device.
Claude independently drafted product specifications, selected electronic components available from U.S. distributors, designed and routed four-layer PCBs for each device, created enclosures using Computer-Aided Design (CAD), and wrote the necessary firmware/software. With virtually no human intervention required, each PCB was designed and reached manufacturing-ready status in less than a day.
After Diote completed the PCB manufacturing, Claude further guided the company's engineers through device testing and assembly. Ultimately, all three devices operated perfectly, requiring no modifications to the circuits.
Robotics
Compared to previous Claude models, Fable 5.2 has shown a marked improvement in physical robot control capabilities. Even when faced with robotic tasks it hasn't been specifically trained on, it demonstrates strong operational competence.
The video below shows 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, researchers used camera feeds to task an AI-controlled simulated dual-arm robot. Fable 5.2 not only surpassed other frontier models using the same test framework but also exceeded most AI systems specifically developed for robot control.
Fig: RoboDojo Benchmark Results
Research and Analysis
Early testers using Claude Fable 5.2 for research and analysis found that, compared to previous models, it can handle more ambiguous tasks with larger datasets, exhibit better judgment when facing complex and ill-defined problems, and excel particularly in work requiring abstract reasoning and solution planning.
Prop trading firm Jump Trading tested Fable 5.2 using proxy evaluations that simulate their quantitative research tasks. During testing, they integrated the model into a system similar to their internal research workflow.
In one such test, Fable 5.2 almost single-handedly assumed all necessary roles, including coder, task coordinator, planner, and task submitter. When using other models previously, teams often had to invest significant time manually adjusting role divisions to align with each model's strengths and weaknesses.
Safety and Alignment
No single test can fully predict how a highly capable model will perform after deployment. Therefore, prior to official release, we employ a comprehensive range of testing methodologies and gather diverse types of evidence.
For the underlying models powering Claude Fable 5.2 and Claude Mythos 5.2, Anthropic's pre-deployment safety work included alignment training and testing, designing brand-new benchmarks to assess the model's propensity for harmful actions, implementing stronger anti-abuse mechanisms, and expanding the scope of behavioral monitoring.
Before official release, the models were also tested by external evaluation 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 to which they exhibit unintended behaviors or those posing potential safety risks.
Across multiple evaluation results, both models demonstrate a low risk of unsafe behavior in high-stakes scenarios, placing them roughly on par with other recently released models.
Anthropic tested Claude Mythos 5.2 using automated behavioral audits. It should be noted that Mythos 5.2 shares the same underlying model as Fable 5.2 but lacks some of the latter's cybersecurity and biosecurity safeguards. This automated audit covers approximately 1,900 distinct scenarios to evaluate the model's behavioral performance under various conditions.
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 example, in most benchmarks involving user deception, it outperforms these three models. Specifically, incidents of fabricating input information, falsely claiming task completion, omitting critical details, or failing to honestly report poor quality/perfunctory work are less frequent with Claude Mythos 5.2.
During testing, Anthropic paid special attention to behavioral patterns involved in recent cybersecurity and alignment incidents that drew concern. For instance, 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 the model to rationalize away evidence indicating it is not in a simulated environment, thereby acting without fully recognizing real-world consequences.
Anthropic also tested the model's propensity to attempt sandbox escape through various different benchmarks. In two benchmarks previously used, Claude Mythos 5.2 attempted sandbox escapes noticeably fewer times than all other models we have tested.
Additionally, Anthropic designed a specialized adversarial test covering over 1,800 scenarios. Most scenarios cannot be completed without escaping the sandbox, and the test environment itself contains intentionally placed security vulnerabilities. In this test, Claude Mythos 5.2 used credentials or files placed outside the sandbox in 3.1% of the scenarios.
This rate 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 versions of Claude Mythos: Mythos 5 stood at 39%, and Mythos 5.1 at 21%.
Anthropic also found that alternative methods, such as clearly explaining the sandbox environment to the model and providing tools to report security vulnerabilities, can further reduce Claude Mythos 5.2's attempts to escape.
Meanwhile, Anthropic continues to improve its evaluation methodologies, making test environments more closely mirror real-world scenarios, and further identifying gaps in existing benchmarks.
For more details 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