Anthropic Accidentally Revealed Their Most Powerful Model Ever

Anthropic's draft blog post on a new model called Claude Mythos leaked from an unsecured database, prompting the company to confirm its existence as their most capable model to date—a step change over Claude 3 Opus in coding, academic reasoning, and cybersecurity benchmarks. The episode also covers Google's real-time voice model, Shopify's e-commerce AI tools, OpenAI updates, and a deep dive into vertical models outperforming generals, exemplified by Intercom's customer service AI.

Anthropic's Claude Mythos Leak

A cache of nearly 3,000 unpublished Anthropic documents surfaced publicly Thursday, including a draft announcing Claude Mythos. Described as far larger and more intelligent than Opus models, it evokes "deep connective tissue" linking knowledge. The post notes dramatically higher scores versus Opus 3.6 but emphasizes caution: Mythos requires extra risk assessment, especially cybersecurity threats, due to its power. As a compute-intensive model, it's costly to run, so Anthropic is optimizing efficiency and starting with early access trials for cyber applications. Possibly codenamed Capybara earlier, no general release timeline was given. Online reactions highlighted the Cthulhu-like name's irony amid AI safety concerns and compared it to OpenAI's potato-named Spud, with speculation it signals accelerating model races.