Anthropic has run into a rare headache: an unreleased version of Claude leaked before its official launch. On June 3 a model identifier codenamed Oceanus showed up in the Claude Console, and the same day access went to vetted red-team specialists — the people who hunt for weaknesses before a launch.
From there it went off script. According to researchers, within hours someone began reselling access to the closed model through a proxy service based in China. In response, Anthropic reportedly paused the testing program for the whole cohort while it looks into what happened.
Oceanus is believed to be part of a coming lineup that leaks call Mythos — a next-generation Claude focused on reasoning, coding, cybersecurity, and autonomous agent work rather than plain chat. A Mythos Preview already appeared in April, and Oceanus looks like the next step.
There's plenty of guesswork around the release. Leakers say a public version could arrive "within weeks," possibly under the name Claude Fable — right around Anthropic's Code with Claude developer event in Tokyo on June 10. One caveat matters here: Anthropic has not officially confirmed Oceanus, its name, or a launch date. All of this comes from third-party reports and leaks, not a company announcement.
Why it's worth watching: red-teaming is the stage where a model is deliberately stress-tested for weak spots before anyone else sees it. A leak at exactly this point is sensitive for both security and reputation — it shows how hard it is to keep a frontier model secret even when only a narrow circle of vetted people has access.
Whether the incident affects the launch timeline is still unknown. But the simple fact that a closed model could be resold within hours is its own signal of how hot the market for access to top AI models has become.
