DeepSeek V4 Pro Scores 83.3 on Cybergym Benchmark

Posts on X report DeepSeek-V4-Pro-0813 reaching 83.3 on Cybergym and 87.9 on Terminal Bench 2.1. Tables place it ahead of some models including Opus-4.8 in several agent metrics while trailing Kimi-K3 in roughly half the evaluations shown. TeortaxesTex suggests the result may reflect V4-Flash-0731 with pass@2 sampling after internal distillation. Other replies note its position above Grok on ProofBench yet below models such as Muse Spark and question further scaling needs.