Enterprise AI is entering a cost-optimization phase as companies shift from using the best models to the most cost-effective ones. Key points: * 95% of enterprise AI workloads still run on premium models * DeepSeek V4 Pro is reportedly 7x cheaper on input tokens and 17x cheaper on output tokens than leading Western models * Enterprises now use a median of 14 AI models to optimize cost and performance * Pinterest reported ~90% cost savings using a customized Qwen model * OpenRouter data shows DeepSeek models processed nearly 6T tokens recently My take: AI is becoming a cost-efficiency game, not just a performance race. Companies are increasingly reserving premium models for complex tasks while routing high-volume workloads to cheaper alternatives. The winners may be infrastructure and routing platforms, not necessarily the most powerful model providers.