Fable 5.1's Terminal-Bench-Science score going from 24.7 to 52.6 percent is the number every outlet led with, and it holds up. The cost curve underneath tells a different story. At the highest effort tier, the model burns close to twice the output tokens to get there, which cuts into the cache savings Anthropic is advertising everywhere else.
 
A quieter change might matter more long-term. Fable 5.1 now locks thinking blocks to the exact context that produced them, breaking a documented technique for training rival models on Claude's reasoning. Paired with a watermark detection API moving into private preview, this release puts real weight behind provenance and copy-protection. We cover the pricing tiers, the Mythos/Fable safeguard split, and what the Life Sciences Verification Program gates.
 
Read the full story: https://literouteracademia.com/2026/09/05/anthropic-ships-claude-fable-5-1-with-a-science-score-that-more-than-doubles/