DeepSeek V4 Pro and Flash split planner/executor roles for token efficiency

Token efficiency in AI coding isn't just about lower-cost models or longer context windows. DeepSeek argues it's an engineering method that divides labor, budgets context, and orchestrates tasks. With V4 Pro and V4 Flash, they have productized the planner/executor split. V4 Pro acts as the consultant: it breaks down requirements, designs architecture, analyzes complex bugs, and performs final acceptance. V4 Flash is the executor: it scans files, writes simple implementations, completes tests, generates candidates, and handles repetitive work. Both models support 1M tokens of context, JSON output, tool calls, chat prefix completion, and FIM completion. Cache-hit input pricing has been reduced to one-tenth of the launch price, making it affordable to repeatedly load prompts, project docs, and history.

This structure prevents wasting frontier reasoning on mechanical tasks. The old approach of using the strongest model for everything is inefficient. Instead, big models (V4 Pro) should appear only at key decision points, while small models (V4 Flash) batch-process and edit. A harness—a scheduling system that splits tasks, validates results, retries failures, and passes context—is essential to keep small models on track. With 1M context and cheap cache hits, DeepSeek enables long-chain agents that can hold entire codebases in memory without forgetting. The result is a realistic cost model for AI-assisted coding, where token consumption directly correlates with verifiable output.