Anthropic's 'dangerous' Mythos class model, Fable 5, is back, and Artificial Lawyer tested it on legal benchmarks and random questions. On High setting, Fable 5 feels assertive and gives answers in multiple long paragraphs. In a self-reported benchmark, Fable 5 stated it is "dangerous in the safety sense" because the underlying model (Mythos 5) has "dual-use capabilities — cyber, bio, chemistry — that Anthropic decided shouldn’t be generally available," so guarded domains fall back to Opus 4.8. Regarding danger to lawyers, Fable 5 reported a benchmark gap of "~90% of individual criteria right, ~11% of complete legal products right," and stated it is "dangerous to the billable hour attached to first drafts, due diligence, research memos" but not yet to judgment or accountability. It added that "the danger to lawyers isn’t that I’m too good — it’s that I’m good enough at 1% of the cost." On Claude for Legal, Fable 5 predicted "real but uneven impact," compressing the total addressable market, with "thin wrappers most exposed" and tools whose moat is UI plus prompting facing "valuation compression within 18–36 months," while workflow depth protects.