DeepSeek V4 Pro Draws Token Backlash

The complaint is plausible at the user-experience layer even if DeepSeek’s architecture is more efficient under the hood: output verbosity and task latency still dominate how “smart” a model feels.

• –DeepSeek’s V4 release says the model is optimized for 1M-context efficiency and lower FLOPs/KV cache than V3.2, so the official story is about backend efficiency, not necessarily shorter answers.
• –If V4-Pro needs more visible tokens to solve the same task, developers pay twice: more latency and more billable output, even when raw inference is cheaper.
• –The 10x comparison to GPT-5-class models is directionally interesting but not a clean benchmark without identical prompts, stop conditions, and task sets.
• –This looks less like a settled regression than a reminder that “intelligence density” is now a product metric, not just a research metric.