SemiAnalysis claims Anthropic's completed Mythos2 is shelved for safety, now used to train Mythos3

新智元 · wechat · 2026-08-18

A SemiAnalysis internal podcast featuring analysts Dylan Patel and Jordan Nanos surfaced multiple unverified claims: Anthropic's new model Mythos2 has finished all training but cannot be released because its capabilities (code generation, reasoning, cyber offense) triggered the highest alarms, making it an internal-only "forbidden" model—while Anthropic now uses Mythos2 to generate data and code to push Mythos3. Patel mocked this as the ultimate boomerang of Anthropic's years of begging for regulation. The most critical consequence may be severing the release-feedback flywheel, but an internal ouroboros flywheel is accelerating: data recursion (strong models generating and verifying their own training data), labor recursion (Claude already writes the vast majority of Anthropic's code, with R&D speedup approaching a 2x warning threshold), and eval recursion (red-teaming and peer review delegated to strong models)—human feedback is demoted from essential to optional, and Anthropic's own evals have "saturated," unable to measure continued acceleration.

On OpenAI, Nanos claimed the rumored 5.6Sol is not the largest model OpenAI has trained—"they definitely have a much bigger model hidden somewhere"; Astra is too strong, with "Critical" cyber capabilities not yet ruled out, so OpenAI paused some internal activities failing new safety requirements. Codex lead Thibault Sottiaux hinted Codex will integrate Astra, and Greg Brockman said OpenAI is building models that write "superhuman-level security code." A popular take speculated Anthropic wants regulation as cover to keep its strongest AI for itself and become a super-app company. All claims are from SemiAnalysis and unconfirmed.

Original post →

SemiAnalysis claims Anthropic's completed Mythos2 is shelved for safety, now used to train Mythos3

More from Models