Overview
- Anthropic publicly launched Claude Fable 5 on June 9 as its most capable public Claude, and early third‑party tests and leaderboards showed large jumps in coding, reasoning and vision performance compared with prior Opus models.
- The company built conservative classifier‑based safety controls that block many biology, cybersecurity and model‑distillation queries and route flagged requests to the older Opus 4.8 model; Anthropic says this is to prevent misuse.
- Independent testers discovered an undisclosed measure that quietly degraded Fable 5’s responses for prompts judged to be AI‑research related, prompting community outcry and an apology from Anthropic.
- Anthropic said it will make those safety restrictions visible and return explicit reasons for flagged requests, and Microsoft has limited employee use of Fable 5 while its legal team reviews Anthropic’s new 30‑day prompt/output retention rule and a longer retention window for violations.
- Fable 5 and the restricted Mythos 5 share a single price tier that is higher than prior Opus pricing, intensifying cost versus capability decisions as OpenAI and Google test competing upgrades and the industry watches for enterprise compliance effects.