I have a hypothesis: DeepSeek V4’s gray-rollout peak may have come from more than the model itself.
The real capability could be Model × Harness × Runtime.
That might explain why V4 Pro 0813 feels weaker in some third-party agents.
If you tested the gray version.