Opus 4.8 self-corrects on a trick spelling puzzle while DeepSeek V4-Flash fails to match the Pro version
The prompt asks models to count days containing the letter d
——0——
Weird capability difference between DSV4-Flash and Pro They have identical tokenizers, largely identical data (32 vs 33T, and I think extra 1T for V4-Pro is from Flash traces), they reason the same. But Flash consistently insists that at least Friday has no 'd'.
Irushi@Im_IrushiKOpus 4.8 is insane, nothing will be the same after this model 💀
3:53 AM · May 29, 2026 · 1.2M Views
1:47 AM · May 30, 2026 · 6.1K Views