Anthropic Releases Claude Fable 5.2 and Mythos 5.2 Flagship Models

At the final juncture before sprinting toward a U.S. IPO, Anthropic has pulled out its latest-generation "most powerful model on Earth."

On Monday, Anthropic announced the release of its new-generation flagship models Claude Fable 5.2 and Claude Mythos 5.2. Judging from the test results the company published, this upgrade is mainly reflected in complex programming, scientific research reasoning, chart understanding and robot control, while output speed is up to 30% faster than the previous generation, and standard input and output prices have both been cut by 20%.

It should be noted that Fable 5.2 and Mythos 5.2 use the same underlying model; the main differences are safety restrictions and access permissions. Fable is open to a broader range of users; Mythos provides access to relevant institutions through a review program, supporting specific cybersecurity and life sciences uses.

This release also means that since Anthropic chief Dario Amodei called in September for slowing the pace of AI development, the company has launched a so-called "frontier" model for the first time. Facing an environment of concern about AI investment, the company must prove it deserves an IPO valuation of $1 trillion, or even $1.2-1.5 trillion.

First, the benchmarks: stronger programming, more accurate chart reading, markedly improved robot control

According to test data published by Anthropic, Fable 5.2 leads the key competitor OpenAI's GPT-6 Astra and Google's Gemini 4 Carbon in almost all of multiple tests, including complex programming, scientific research tasks, computer operation, robot control, interdisciplinary reasoning and chart recognition.

In the FrontierCode 2.0 test, which measures complex programming capability, Fable 5.2's score rose from 32.4% in the previous generation to 49.1%, higher than Opus 5.5's 45.5%, GPT-6 Astra's 42.8% and Gemini 4 Carbon's 41.1%.

This test focuses on whether the code changes a model makes to open-source projects meet the standard that project maintainers are willing to accept. In other words, it tests whether AI can write code that is truly usable and suitable for existing projects. The new model improved by 16.7 percentage points over the previous generation, reflecting progress in handling complex software engineering problems.

In the chart understanding test Chartography, Fable 5.2's score rose from 44.8% to 73.4%. This type of capability concerns whether a model can accurately read charts, coordinates and data relationships, and is the foundation for analyzing financial report charts, experimental results and business reports.

In terms of robotics, Fable 5.2 achieved a success rate of 64.9% in the RoboDojo simulated dual-arm robot test, more than double that of GPT-6 Astra. This improvement means the model's ability to judge the environment from camera images and plan actions has clearly strengthened.

Scientific research reasoning has also improved. In the "Humanity's Last Exam," which allows tool use, Fable 5.2's score rose from 65.6% to 74.2%. This test mainly examines interdisciplinary high-difficulty knowledge and reasoning, reflecting the model's ability to handle specialized problems.

Easier to understand than test scores is what the new model can actually do.

Anthropic disclosed that an engineer once asked Fable 5.2 to develop an operating system prototype. Starting from an empty folder, the model orchestrated more than 1,000 sub-agents and spent eight consecutive days writing, testing and reviewing code, ultimately producing an operating system that could boot on a laptop.

This prototype system not only has windows and a login interface, but also an operating system kernel responsible for managing hardware resources and scheduling program execution, with Claude functionality built in.

Another case closer to real enterprise applications is that Anthropic used a team of agents running an early version of Fable 5.2 to migrate a key production service from Python to Rust. The company said that after the migration the service is more efficient and reliable, with startup speed about 26 times the original.

For enterprises, the significance of this type of capability is that some system overhaul projects that were shelved because they took too long or cost too much may begin to become worth implementing.

Completing circuit board design within a day

Anthropic partnered with circuit board design and manufacturing company Diode to have Fable 5.2 design three devices based on a brief set of requirements: a home voice assistant hub, a USB conference speaker and a handheld virtual pet.

The model was responsible for drawing up specifications, selecting electronic components that could be purchased, designing the routing for a four-layer circuit board, drawing the enclosure and writing the software. The company said that with a small amount of human intervention, each circuit board reached a state ready for manufacturing delivery within a day.

Subsequent manufacturing, testing and assembly were completed by Diode engineers, with Claude providing guidance. In the end, all three devices ran successfully without any circuit modifications.

On robotics, the company also showed footage of Fable 5.2 controlling robots to complete tasks such as unscrewing bottle caps, manipulating a Rubik's cube and drawing. The company said that compared with earlier models, Fable 5.2 has significantly improved in controlling physical robots, even in robot tasks for which it had never received specialized training.

There are also cases in theoretical physics and biopharmaceuticals

In theoretical physics and mathematics, Anthropic also presented three results from the new model: solving two theoretical physics problems related to black holes, and completing the formalization of the proof of Fermat's Last Theorem.

First, the model completed a complex calculation originating from a 2003 study. Put simply, the researchers wanted to know: in a specific theoretical model, does a sudden change related to black holes, akin to "water boiling," still exist under conditions that require description by quantum theory? The company says the model computed the key result, and the answer is yes.

Second, the model determined the capability ceiling of a class of mathematical methods. This method is used to screen candidate theories of quantum gravity and rule out schemes that do not meet the criteria. This result is equivalent to calculating the finest level this mathematical "sieve" can achieve, helping researchers judge how much further existing methods can go. Preprints of both physics studies have been published.

Finally, about 1,000 Fable 5.2 agents collaborated to complete the formalization of the proof of Fermat's Last Theorem in 30 hours. This is not the first proof of Fermat's Last Theorem, but rather the construction of a complete proof that a computer can verify step by step, requiring the filling in of derivation steps and the coordination of the logical relationships among the parts. Under the same thousand-agent configuration, Opus 5.5 achieved a weighted completion rate of about 27% after running for 30 hours. However, this comparison tested each model only once.

In the life sciences, Mythos 5.2 also developed a method for predicting the complete structure of viral capsids based on existing open-source protein folding models. The capsid is the protein shell of a virus and also a delivery vehicle used in some gene therapies. The company says that of the 34 capsids attempted, the method accurately modeled 28.

Mohammed AlQuraishi, Assistant Professor of Systems Biology and Computer Science at Columbia University, said that large protein complexes, including viral capsids, have long been a difficult problem in the field of protein modeling: one must accurately predict the structures of dozens of macromolecules and their relative positions, while facing extremely high computational costs. Claude's method cleverly addresses the modeling challenge of symmetric complexes, giving researchers hope of completing on a personal workstation work that previously required a dedicated computing cluster.

In research and analysis, the progress demonstrated by Fable 5.2 includes both processing data and writing code, as well as discovering contradictions in materials and judging what conclusions the evidence can support.

Quantitative trading firm Jump Trading tested the model's performance in a process similar to internal research workflows. Fable 5.2 was able to take on multiple roles such as posing questions, writing code, making plans and coordinating tasks, reducing the burden of manually adjusting the division of labor. The company says that in an internal research automation test, the model's error was 10% lower than that of the second-ranked model.

Investment firm Balyasny Asset Management had the model investigate how an institution tests new integration features in core technical systems. According to its feedback, Fable 5.2 found the relevant sources, did not overstate what the evidence could prove, and also pointed out that the wording of the question itself might be misleading. For investment research, this type of capability helps distinguish established facts from judgments still awaiting verification.

Trading firm IMC focused more on whether the model could propose new research ideas. The company said that Fable 5.2 not only reproduced research insights that had previously impressed it, but also proposed an entirely new angle, and achieved new best results on three of the four quantitative research questions that had completed testing.

In a financial modeling task, Fundamental Research Labs asked the model to extract data from a company's 10-K annual report and fill it into a forecasting model. Faced with a situation where the template did not clearly specify whether a certain metric should use the disclosed value or a calculated value, Fable 5.2 judged the requirement based on cell formatting, and also discovered that the sum of regional data was inconsistent with the disclosed total, ultimately correctly filling in all target cells.

Price cut by 20%, strengthening commercial competitiveness before launch

Beyond capability improvements, Anthropic is also lowering the cost of using its flagship model.

The new model's standard price is $8 per million input tokens and $40 per million output tokens, both down 20% from the previous generation. The company also launched an optional fast mode, in which token generation speed can be increased by up to 2.5 times, with input and output prices of $16 and $80, respectively.

Anthropic emphasizes that for most everyday programming and knowledge tasks, users can still start with Opus 5.5, and use the flagship model when facing complex, open-ended problems.

(Source: Cailian Press) Middle of the article: embedded advertisement