A hands-on test of V4-Pro's default reasoning mode
Tested on 2026-08-14 using deepseek-v4-pro on DeepSeek's official endpoint (the production V4-Pro-0813 release).
TL;DR
Reasoning is on by default. With no reasoning-related parameter, a single Liar's Dice decision consumed 3,072 output tokens—all reasoning, with no visible answer at all.
The OpenAI-style fix did nothing. I sent max_completion_tokens: 3072. The API returned HTTP 200, then generated 15,809 tokens over 222 seconds. It did not reject the parameter; it simply behaved as if it had never seen it.
