Anthropic implemented new safeguards restricting access to Claude's reasoning that reduced reasoning tokens extracted from Claude AI models by 90% to 99%, or 10 to 100 times lower than previous levels