Anthropic releases Fable and Mythos 5.2 models before IPO

New flagship models for technical tasks

Anthropic has launched two new flagship models, Claude Fable 5.2 and Claude Mythos 5.2, aimed at the most demanding technical requirements. These models are built on the same foundation but are distinguished by the level of access and safety restrictions applied to them. While Opus 5.5 remains the company's recommendation for everyday use, these new models target specialized high-end work.

Fable 5.2 is available to all users and includes safeguards that block risky tasks in biology and cybersecurity. In contrast, Mythos 5.2 removes some of these restrictions but is only available to organizations that apply through specific verification programs. This dual-model approach allows Anthropic to provide high-performance tools while maintaining safety boundaries for the general public.

The cost for these models is set at $8 for every million input tokens and $40 for every million output tokens. This represents a price cut from the previous versions, which cost $10 and $50 respectively. However, the actual cost for a user will depend on specific token consumption, which fluctuates based on the software environment and the complexity of the task performed.

Benchmark leads in intelligence

The Fable 5.2 model has taken a clear lead on several industry benchmarks. According to Artificial Analysis, it scored 65 points on the Intelligence Index v5, placing it five points ahead of the unreleased Gemini 4 Carbon. It also outperformed its predecessor, Fable 5.1, by 15 points in the same evaluation.

The model demonstrated significant strength in academic and scientific workflows. It achieved an 80 percent score on HLE-Diamond, a test featuring expert-level academic questions. Additionally, it scored 73 percent on Terminal-Bench-Science, which measures the ability of an AI agent to run scientific workflows independently in a terminal environment.

While Fable 5.2 excels in these areas, it shows some relative weaknesses in multi-step knowledge work and terminal-based coding. It reached 1,785 Elo on the AA-Briefcase benchmark, trailing slightly behind Claude Sonnet 5.5. On Terminal-Bench 4.0, it tied with Opus 5.5 at 60 percent, while other models like Kimi K3.1 Pro ranked higher in that specific category.

Advancements in cybersecurity

The Mythos 5.2 model has set new records for finding security flaws in software. In a test using the CVE database, which catalogs known software vulnerabilities, Mythos 5.2 identified 10.2 percent more flaws than Opus 5.5. It also outperformed the previous Mythos 5.1 model by 14.4 percent during the October 2026 run involving 1,829 CVEs.

Anthropic reported that one in six of the confirmed bugs found by Mythos 5.2 had been missed by all other frontier models tested. The model is also noted for being better at reproducing and fixing these vulnerabilities once they are identified. Because of its high capabilities, the publicly available Fable 5.2 variant blocks most security-related tasks like penetration tests.

When a user submits a blocked cybersecurity request to Fable 5.2, the system redirects the request to Opus 4.8. Similarly, requests involving biology are directed to Opus 5. This routing system ensures that the most powerful capabilities remain under the control of verified organizations. These findings come amid intense ongoing debates regarding the security implications of frontier AI models.

Formalizing complex mathematics

One of the most notable achievements for the new models involves the formal verification of Fermat's Last Theorem. About 1,000 Fable 5.2 agents successfully translated the proof into the Lean programming language. This allows a computer to verify the proof step by step, a feat that had not been completed for machine verification until now.

The process took 30 hours to complete using the new agents. This is a significant improvement over an earlier, unpublished research version of Claude that required 11 days and human guidance. The Fable 5.2 agents were able to organize their own collaboration and produced a proof that was seven times shorter than the first attempt.

Anthropic compared these results against Opus 5.5, which managed only 26 percent of the formalization in 18 hours. The Opus model made almost no progress after the 18-hour mark. This highlights the specialized ability of the Fable 5.2 agents to handle complex, multi-step mathematical logic without relying on external mathematical references.

Safety and development limits

Anthropic stated that it deliberately prevented the new models from gaining certain capabilities during the training process. The company focused on avoiding improvements in areas that could directly accelerate AI development, a concept known as RSI. The models match the performance of Opus 5.5 when it comes to tasks like refining compute kernels for chips or improving the structure of large language models.

The company clarified that these limitations do not affect other related capabilities. This includes agentic coding, hardware development, and traditional AI research such as interpretability research. Despite these restrictions, the models remain highly capable in many other domains. For instance, Claude currently generates more than 90 percent of the code updates merged into the company's main codebase while being guided by humans.

In internal testing, the model even demonstrated the ability to program an operating system from scratch. These releases come just before Anthropic's planned initial public offering. By balancing high-performance benchmarks with deliberate training constraints, the company aims to maintain a lead in the AI market while addressing concerns about rapid development acceleration.