QUICK VERDICT — POST JULY 24 MIGRATION
● Cheapest option: V4 Flash — $0.14/$0.28/M (3x cheaper than V4 Pro)
● Strongest reasoning: V4 Pro — $0.44/$0.87/M
● Thinking mode: Both default ON — set OFF explicitly for fast tasks
● Context: Both 1M tokens
● Migration trap: deepseek-reasoner maps to Flash by default — if you needed Pro quality, update your model string
| Model | Input /1M | Output /1M | Thinking | Context | Best for |
|---|
| deepseek-v4-flash | $0.14 | $0.28 | ON (set OFF for speed) | 1M | High-volume, routing, classification |
| deepseek-v4-pro | $0.44 | $0.87 | ON (set OFF for speed) | 1M | Complex reasoning, hard coding tasks |
V4 Flash for: High-volume text generation, routing and classification pipelines, document processing, agentic fan-out subagents, and any workload where $0.14/M input is the deciding factor. Set thinking OFF for latency-sensitive tasks. Was the default replacement for deepseek-chat.
V4 Pro for: Complex multi-step reasoning, hard coding tasks, research synthesis, and any workload where quality is the constraint rather than cost. 3x more expensive than Flash but still 5-10x cheaper than Western frontier alternatives. Must be called explicitly — does not appear as a default migration target.
The thinking mode decision: Both default thinking ON. For reasoning tasks (math, code, complex analysis) — leave it ON. For fast generation tasks (summarisation, translation, classification) — set it OFF. Not doing this will add unnecessary latency and tokens to every fast task call.
Last updated July 24, 2026. Related: DeepSeek V4 migration guide → · DeepSeek V4 Pro vs Gemini 3.6 Flash →