antirez called out where Fable's capability jump shows up most clearly: "Especially: Fable is impressive in doing this that are very under-represented in its dataset. The model is generally much better to reason about any kind of code, but it is kinda impressive how you can point it to something very niche and see it delivering." The distinction: general code reasoning improved across the board, but the standout result is on niche, low-frequency-in-training-data problems, the cases where a model has the least statistical pattern to lean on and has to actually reason instead of pattern-match. That gap, more than aggregate benchmark gains, is his marker of a real capability increment. **Cross-Domain Connections**: [[Fable Makes Previously Impossible Expert-Level Programming Conversations Possible]] is the broader claim this specifies. [[Explore an AI Model's Capability Ceiling With Maximal Requests First]] is a complementary practice for surfacing this kind of capability: probe with the most demanding, least-represented request first. ## Source - [[Antirez Fable Feedback]] — X (Twitter) thread, @antirez, June 11, 2026