Abstract:
Anthropic officially launched its new generation flagship artificial intelligence models Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The new version focuses on improving programming, logical reasoning and visual understanding capabilities, while improving the expression quality of the model. According to data released by Anthropic, the output speed of the two models is up to 30% faster than the previous generation, while the API call price is reduced by 20%.
Fable and Mythos are the highest-positioned model series in Anthropic's product system, mainly oriented to complex, open technical tasks. In contrast, Haiku, Sonnet and Opus respectively undertake daily tasks of varying scale and complexity. Anthropic recommends that general programming and knowledge work can usually use Claude Opus 5.5 first, and only when faced with particularly complex and time-consuming tasks, you need to switch to Fable or Mythos.
In terms of performance, Anthropic said that Fable 5.2 and Mythos 5.2 are not only able to generate content faster, but are also better at expressing ideas clearly when dealing with complex tasks. Fable 5.2 can also use fewer Tokens than Fable 5.1 when completing difficult tasks under moderate inference intensity, thus further reducing actual usage costs.
Benchmark tests published by Anthropic show that Fable 5.2 surpasses its own Opus 5.5 and the previous generation Fable 5.1 in multiple tests. In the FrontierCode 2.0 programming test, Fable 5.2 achieved a score of 49.1% at the highest inference intensity, which was significantly higher than Fable 5.1's 32.4% and also exceeded Opus 5.5's 45.5%. In the FrontierSWE v2 test, the new model scored 69.9%, compared to 56.3% for Fable 5.1 and 62.3% for Opus 5.5.
In the Humanity's Last Exam test involving complex knowledge reasoning, Fable 5.2 achieved a score of 74.2% when used with tools, which was higher than Fable 5.1's 65.6% and Opus 5.5's 67.7%. In the RoboDojo Sim dual-arm robot simulation test, the new model scored 64.9%, a significant improvement compared to the previous generation's 27.6%. Additionally, in the Remote Labor Index test, the Fable 5.2 scored 41.3%, while the Fable 5.1 only scored 17.9%.
However, the new model did not lead in all tests. For example, in the AutomationBench enterprise workflow test, Google Gemini 4 Carbon achieved 53.9%, higher than Fable 5.2's 51.4%. These results also show that there are still differences in the performance of different models on specific task types, and comprehensive capabilities cannot be judged based on a single benchmark test alone.
Anthropic also introduced the actual performance of Fable 5.2 in internal software development. The company said that currently more than 90% of the merged code modifications in its core code base have been completed by Claude under manual guidance. The engineering team used Fable 5.2 to fix long-standing software defects, optimize the operating efficiency of multiple services, and complete some projects that had been difficult to launch in the past due to cost or time constraints.
One case is particularly interesting: a group of AI agents using an early version of Fable 5.2 migrated a key production service from Python to Rust. After the migration was completed, the service started approximately 26 times faster, while operating efficiency and reliability were also improved. Anthropic believes that this reflects the new model's progress in engineering judgment and handling ambiguous requirements.
This release also involves adjustments to the artificial intelligence security mechanism. Fable 5.2 and Mythos 5.2 actually use the same underlying model, but have different security restrictions configured for different user groups. The publicly available Fable 5.2 sets up security protection mechanisms in the fields of cybersecurity and biology to block high-risk requests that may be abused. Anthropic said that compared with Fable 5.1, the new version reduced the number of security interceptions in early tests by about 55%, meaning the model is more flexible in handling normal requests while still retaining restrictions for high-risk uses.
Mythos 5.2 is open to audited professional organizations, mainly related to network security and life science research. Because its security restrictions are different from the public version, ordinary users cannot directly access Mythos like they can with Fable. Anthropic hopes that through the review mechanism, it will allow professional researchers to carry out difficult work while reducing the risk of technology abuse.
In terms of price, this upgrade also brought significant changes. The API prices of Fable 5.2 and Mythos 5.2 are both US$8 per million input tokens and US$40 per million output tokens, compared with US$10 and US$50 respectively for the previous generation models, a decrease of 20%. In addition, Anthropic also provides an optional fast mode, which can increase the token generation speed up to 2.5 times that of the normal mode, but the corresponding price will also rise to US$16 per million input tokens and US$80 per million output tokens.
Fable 5.2 has been opened to Claude paid users and provides services through API. It also supports platforms such as AWS, Google Cloud and Microsoft Azure. Anthropic also announced that eligible paying subscribers will receive additional usage credits this week, including one to three usage credit resets, depending on the length of the subscription. Users of Pro plan and Team plan standard seats can also use up to 50% of their weekly usage quota for the Fable model.
Overall, the Fable 5.2 upgrade is not just a regular model performance iteration. Faster generation speed, lower call cost, and improved performance in complex programming and multi-step agent tasks mean that Anthropic is further strengthening Claude's competitiveness in professional technical work. Mythos 5.2 continues to adopt a restricted open strategy, which also reflects the company's product philosophy of combining the strongest capabilities with security review.
It should be noted that the above benchmark results and internal application cases are all from information released by Anthropic and cannot be completely equivalent to the test results of independent third parties in all actual usage scenarios.
Comments