High cost of cybersecurity: Anthropic Mythos is five times more efficient than traditional tools, burning millions of tokens in three weeks According to Beating monitoring, in the early enterprise testing of Mythos, a large model developed by Anthropic specifically for network security, security giant Palo Alto Networks found that although it identified more than 20 key vulnerabilities in the company's source code within three weeks, with a detection efficiency five times that of traditional tools, it consumed more than $1 million in token quota in a short period of time. Another testing company also confirmed that using only two weeks of book costs could potentially reach millions of dollars. The high cost is mainly due to high pricing. According to Anthropic's official pricing, the price per token for Mythos' API is approximately 6 times that of the flagship model Opus. Due to concerns about hackers potentially abusing tools to assist in cyber attacks, Anthropic has long restricted public access to Mythos and only opened testing to a few enterprises. However, as the evaluation by the UK AI security testing agency shows that Mythos' performance is slightly better than OpenAI's yet to be released competitive model GPT-5.5-cyber, Anthropic has decided to fully open its API interface to all customers in the coming weeks. To control expenses, early testing companies are optimizing large model scheduling through engineering methods. Enterprise software vendor UiPath has redesigned its system prompt words, reducing the consumption of warm tokens from 150000 to 3000 per task startup. Palo Alto Networks has developed a master-slave dual model collaborative architecture, with the high cost Mythos responsible for developing analysis path planning for intrusion and vulnerability scanning, and then assigning specific code execution tasks to the lower cost Opus 4.7. The master-slave dual model collaboration enables enterprises to enjoy top-notch security performance while successfully suppressing computing power expenditures. [Original link]