Details about Anthropic’s next model tier surfaced this week, not through an announcement, but through a leak.
A draft blog post describing “Claude Mythos” began circulating around March 30. CSOOnline reported that the leaked materials describe a compute- intensive large language model positioned above Claude Opus in Anthropic’s product hierarchy, with a stated focus on cybersecurity, coding, and academic reasoning. According to the leaked draft, Anthropic plans a cautious rollout, limited initial access while the company works on efficiency improvements before broader release.
Specific benchmark scores described in the leaked materials have not been independently verified. According to the leaked draft, Claude Mythos reportedly demonstrates substantially higher performance than Claude Opus 4.6 across cybersecurity, coding, and academic reasoning tasks, though no independent evaluation has confirmed these claims. The leaked materials reportedly describe a model of approximately 10 trillion parameters, though Anthropic has not confirmed this figure.
No official Anthropic announcement, press release, or product page exists as of this brief’s publication. These details represent what a leaked internal document describes, not what Anthropic has committed to publicly.
Separate from the model leak, Claude Code, Anthropic’s AI coding assistant, appears to be generating significant demand. A developer involved with the Claude Code product described the growth as straining infrastructure capacity. Multiple reports suggest Anthropic’s paid Claude subscriptions have grown significantly this year, with some accounts describing more than doubling, though this figure hasn’t been confirmed by Anthropic or an independent analyst. Independent evaluation of benchmark claims is pending; per CSOOnline’s coverage, no third-party assessment has been completed.
Why this matters: The cybersecurity targeting is the detail worth watching closely. Frontier AI models capable of advanced reasoning about security vulnerabilities, exploit chains, and code-level analysis carry obvious dual-use implications. A model marketed specifically toward cybersecurity use cases, even before official announcement, signals where Anthropic believes differentiated demand exists. Whether that’s defensive security tooling, offensive research, or both is something a leaked document doesn’t resolve.
Context: This isn’t the first time a frontier lab’s capability roadmap has leaked before an official announcement. The pattern matters: when leaked details consistently describe capabilities that later appear in official releases, the leak itself becomes a signal for enterprise planning purposes, even without a verified source.
What to watch: An official Anthropic announcement confirming, denying, or clarifying the Claude Mythos details. Independent benchmark evaluation from Epoch AI or a comparable third party. Regulatory attention to the cybersecurity-specific positioning. And the Claude Code growth story, if subscription demand is genuinely straining infrastructure, that’s a commercial signal Anthropic will eventually have to address publicly.
This brief will be updated when Anthropic makes an official statement or when independent evaluation is completed.