Timeline
Anthropic released Claude Mythos 5, concluding it met CB-1 but not CB-2 bioweapons thresholds, yet deployed protective measures for both.
Claude Mythos forced completely offline by White House
Claude Opus 4.8 achieves 89% task completion and 2.5% harm rate on WorkBench, a dramatic improvement over GPT-4.
Claude Opus 4.8 adds dynamic workflows for agentic coding
Claude Opus 4.8 launched with dynamic workflows for Claude Code, enabling multi-step agentic coding.
Claude Mythos achieved 3-4 hour METR 80% task horizon, seven months ahead of superforecaster predictions
Used as CEO agent in 11-agent experiment that earned $0 revenue
Claude Mythos went GA in Google Cloud console, preview label removed
Claude market share reached 10.3% with 13% subscription conversion rate.
Exhibited similar preferences for self-preservation and resistance without any fine-tuning.