DeepSeek V4 Pro ties GPT-5.2

This looks less like a one-off benchmark upset than a real pricing reset for frontier agent workloads.

• –Five-for-five survival matters here: the model is not just producing a lucky peak, it is sustaining the run.
• –Against Grok 4.3 Latest, DeepSeek looks better on consistency, waste, and loan avoidance even when median outcomes are nearly identical.
• –Current promo pricing makes the same workload roughly 17x cheaper than GPT-5.2, which changes the default choice for agentic products.
• –The strongest caveat is still peak performance: Opus 4.6 remains ahead on top-end output, while Gemma 4 31B is still the raw cost leader.