TL;DR — Key Takeaways
- Anthropic released Claude Fable 5.2 and Mythos 5.2 with API prices reduced by 20% and output speeds up to 30% faster than their predecessors.
- Fable 5.2 scored 49.1% on a coding benchmark and coordinated more than 1,000 AI agents to build a prototype operating system in an Anthropic experiment.
- Anthropic reported improved vulnerability detection with Mythos 5.2, which offers specialized cybersecurity and biological research capabilities through verification programs.
Anthropic has given its latest Claude models stronger coding capabilities while making them faster and cheaper to run. The company released Claude Fable 5.2 and Mythos 5.2 on Monday, touting improvements in software engineering, scientific research and cybersecurity.
Anthropic has reduced API prices for both models by 20%, bringing costs to $8 per million input tokens and $40 per million output tokens. The company said the models also generate output up to 30% faster than their predecessors. Developers can choose an even faster processing option, although it costs twice the standard API rate.
In testing reported by Anthropic, Fable 5.2 scored 49.1% on FrontierCode 2.0 at its highest reasoning setting, compared with 32.4% for Fable 5.1. Developed by AI coding company Cognition, the benchmark evaluates AI agents on demanding, real-world software development tasks. Anthropic said the new model achieved the highest score among the publicly released models it tested.
Anthropic also tested Fable 5.2’s coding capabilities in a separate experiment in which an engineer tasked it with building an operating system from scratch. Over eight days, the model coordinated more than 1,000 AI agents to write, test and review code, producing a working prototype that boots on a laptop, complete with windows, a login screen, file storage and hardware drivers.
Cybersecurity is another area where Anthropic reported significant gains. In the company’s internal CVE Rediscovery evaluation, which measures how well models can independently discover known security flaws in older software versions, Mythos 5.2 identified 14.4% more previously disclosed software vulnerabilities than Mythos 5.1. Mozilla distinguished engineer Brian Grinstead also reported that Mythos 5.2 roughly doubled the vulnerability discovery results of its predecessor in the organization’s testing.
Fable 5.2 and Mythos 5.2 use the same underlying AI model, but Anthropic applies different safety restrictions to each. Though Fable is generally available, it includes safeguards that limit certain cybersecurity and biological research activities that could be misused. Anthropic provides access to Mythos through separate Cyber and Life Sciences Verification Programs, which screen users applying to conduct specialized research. The cybersecurity program has three permission levels covering defensive security work, authorized penetration testing and higher-risk research.
Anthropic also reported that Fable 5.2 solved two longstanding theoretical physics problems involving black holes, including a calculation proposed in 2003 about how black holes form in very small theoretical universes. In biological research, the company said Mythos 5.2 developed a method for predicting the structures of viral protein capsids used to deliver gene therapies, accurately modeling 28 of 34 in its tests.
Fable 5.2 is available through Anthropic’s API and cloud platforms including AWS, Google Cloud and Microsoft Azure. Anthropic said Claude Enterprise customers can also use Mythos 5.2 to scan code for vulnerabilities through Claude Security without joining the Cyber Verification Program.
