What's this about?
A week after launch, OrangeAI's conclusion on DeepSeek V4 Pro is blunt: it was trained badly. Two pieces of evidence. First, the much cheaper V4 Flash outperforms Pro. Second, and more damning, V4 Pro at High reasoning effort beats V4 Pro at Max. He allows that Pro still writes with better style, while calling the gap to Opus 4.6 "night and day." karminski3 published a video test asking the same questions directly, including whether pairing it with DeepSeek Harness could rescue the release.
What exactly does "High beats Max" tell you?
Reasoning effort is a dial: higher settings let the model think longer before answering, and more thinking should not make a model worse. When a higher setting produces worse results, it usually means the extended reasoning is degrading rather than refining — the model talks itself out of correct answers, or the training rewarded a pattern that doesn't hold at longer lengths. That's a training pathology rather than a tuning preference, which is why this specific inversion is treated as evidence of a bad run rather than a debatable trade-off.
Why is this worth noticing?
- It's a rare public post-mortem of a frontier-adjacent release going wrong, in a discourse where launches are covered and follow-ups usually aren't.
- It complicates the week's other DeepSeek story — the price rises that hit downstream resellers — since the flagship justifying those prices appears to underperform its own cheaper sibling.
- The "does the harness save it?" framing is telling. That harness scaling is now the first thing people reach for when a model disappoints says a lot about where the field thinks capability lives.