In early enterprise trials of Anthropic's cybersecurity-focused large language model Mythos, security giant Palo Alto Networks discovered that while it found over 20 critical vulnerabilities in the company's source code within three weeks — detecting bugs at five times the efficiency of traditional tools — it also consumed more than $1 million in token credits. Another testing company confirmed that just two weeks of use could rack up costs in the millions.
The high cost stems largely from pricing. According to Anthropic's official rates, Mythos's API costs about six times more per token than its flagship model Opus. Due to concerns that hackers might abuse the tool for cyberattacks, Anthropic had long restricted public access to Mythos, limiting it to a handful of enterprise testers. However, after the UK's AI safety testing agency found that Mythos slightly outperforms OpenAI's still-unreleased rival model GPT-5.5-cyber, Anthropic has decided to open the API to all customers in the coming weeks.
To rein in costs, early testers are engineering their model pipelines. Enterprise software company UiPath redesigned its system prompts, cutting warm-up token consumption from 150,000 to just 3,000 per task. Palo Alto Networks developed a master-slave dual-model architecture: the expensive Mythos handles intrusion and vulnerability scan planning, while cheaper Opus 4.7 executes the actual code work. This hybrid approach lets companies enjoy top-tier security performance while keeping compute costs under control.