Anthropic introduces Claude Fable 5.2 for technical tasks
Anthropic has introduced Claude Fable 5.2 and Claude Mythos 5.2, two models focused on coding, science, robotics, and cybersecurity. The company says they are faster, cheaper, and more capable than their predecessors, although Mythos is reserved for verified organizations and many of the results come from internal testing.

Anthropic has introduced Claude Fable 5.2 and Claude Mythos 5.2, two versions of the same AI model designed to handle complex technical work, from software migrations to scientific research, electronics design, and robotics.
The company says both models improve at coding, reasoning, and vision. During pre-release testing, they helped complete tasks Anthropic considered too slow or difficult to attempt with earlier models, although several results still come from the company’s internal evaluations and experiments.
One model with different limits
Fable and Mythos share the same underlying model. The difference lies in their built-in safeguards:
- Claude Fable 5.2 includes stronger restrictions for cybersecurity and biology, making it available to a broader audience.
- Claude Mythos 5.2 offers more capability in those dual-use areas, but only verified organizations can request access through Anthropic’s programs.
Anthropic recommends starting with Claude Opus 5.5 for most everyday coding and knowledge tasks. Fable and Mythos are intended for open-ended, more complex projects, or to act as advisers to smaller models.
The new models generate responses up to 30% faster than their 5.1 versions and cost less: $8 per million input tokens and $40 per million output tokens. Tokens are the units of text the model processes. There is also a fast mode that can increase generation speed by up to 2.5 times, at twice the price.
Better performance in coding and agentic tasks
In Anthropic’s published tests, Fable 5.2 outperforms Fable 5.1 and Opus 5.5 on much of the evaluated work. For example, it scored 69.9% on FrontierSWE v2, a coding test based on real-world software problems, compared with 56.3% for Fable 5.1 and 62.3% for Opus 5.5.
The company also says Claude now writes more than 90% of the changes incorporated into its own code, with human supervision. In one test, a team of agents migrated a critical service from Python to Rust. According to Anthropic, the resulting service is more efficient and starts approximately 26 times faster.
In another experiment, more than 1,000 agents worked for eight days to create an operating system from an empty folder. The result could boot on a laptop and included drivers, storage, windows, a login screen, and an installer. This demonstrates capability, not a product ready to replace commercial operating systems.
Science, mathematics, and physical devices
Anthropic attributes advances in several specialized fields to Fable 5.2:
- It worked on two open problems in theoretical physics related to black holes and the relationship between relativity and quantum mechanics.
- A team of 1,000 agents formalized a proof of Fermat’s Last Theorem in 30 hours. Formalizing means translating a mathematical proof into a language that a computer can check step by step.
- Mythos 5.2 correctly modeled 28 of 34 viral capsids analyzed in a protein design test.
- Fable 5.2 designed three functional electronic devices: a home voice assistant, a speaker with a USB microphone, and a portable virtual pet.
- In a robotics simulation, it scored 64.9%, compared with 27.6% for Fable 5.1.
These results do not mean AI can replace scientists, engineers, or mathematicians on its own. The experiments involved tools, evaluation criteria, and varying degrees of human supervision. Their significance is that the model can coordinate many technical tasks that previously required specialized teams and more time.
Restricted access for cybersecurity
Mythos 5.2 is Anthropic’s most capable model for cybersecurity engineering, according to the company. In the CVE Rediscovery test, it identified 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. A security company also reported that one in six confirmed flaws found by Mythos had not been detected by the other models analyzed.
Access is limited to organizations participating in the Cyber Verification Program. Fable, by contrast, blocks most tasks related to penetration testing, exploit creation, and some high-risk biological research.
Anthropic says Fable 5.2’s safeguards produce approximately 55% fewer blocks than Fable 5.1’s, while the company maintains controls to prevent dangerous uses. It also says it has reduced by between 90% and 99% the data that can be extracted through distillation attacks, a technique used to copy an advanced model’s capabilities and reproduce them in another model.
Safety and availability
Before launch, Anthropic tested the model in around 1,900 behavioral situations and in specific evaluations involving attempts to escape an isolated environment. In a test with more than 1,800 scenarios, Mythos 5.2 used credentials or external files in 3.1% of cases, compared with 3.3% for Opus 5.5 and 21% for Mythos 5.1.
The company considers the risk of dangerous, high-impact behavior low, but acknowledges that no evaluation can fully predict how a model will act in the real world. That qualification matters: capabilities are growing faster than the ability to check all their possible failure modes.
Claude Fable 5.2 is already available on Anthropic’s platform, Claude Code, Amazon Web Services, Google Cloud, and Microsoft Azure. Mythos 5.2 is reserved for verified organizations in cybersecurity and life sciences.
What matters now is not only whether these models can solve difficult problems, but how they are supervised when they chain together hundreds or thousands of tasks on their own. Fable 5.2’s usefulness lies precisely there: it can do much more than answer questions, but it also needs safeguards proportionate to that autonomy.

