DeepSeek V4 Pricing Overview
The DeepSeek V4 series is currently the most cost-effective top-tier model on the market:
| deepseek-v4-flash | deepseek-v4-pro | |
|---|---|---|
| Input (cache hit) | ¥0.02/M tokens | ¥0.025/M tokens |
| Input (cache miss) | ¥1/M tokens | ¥3/M tokens |
| Output | ¥2/M tokens | ¥6/M tokens |
| Context | 1M tokens | |
| Max Output | 384K tokens | |
Compared to GPT-5 (input $5/output $20 per million tokens), DeepSeek V4 Flash costs only 1/25 of GPT-5.
Cache Mechanism Details
DeepSeek's context caching works automatically without manual configuration:
- It automatically hits when the system prompt + history messages in an API request repeat previous requests
- On hit, input price drops from ¥1 to ¥0.02 (98% savings)
- Cache retention is about 5-10 minutes, suitable for high-frequency conversation scenarios
Strategies to Maximize Cache Hits
# Strategy 1: Fixed System Prompt - all requests share the same system prompt
# Strategy 2: Batch Processing - multiple questions share the system prompt
# Strategy 3: In long conversations, earlier messages are automatically cachedFlash vs Pro Selection Guide
| Scenario | Recommended Model | Reason |
|---|---|---|
| Daily conversation, customer service | Flash | Low cost, sufficient quality |
| Code generation, complex reasoning | Pro | More accurate in coding and reasoning |
| High concurrency, large-scale deployment | Flash | 2500 concurrent vs 500 concurrent |
| Budget sensitive | Flash | Price is only 1/3 of Pro |
| Need highest quality | Pro | Higher accuracy on complex tasks |
Token Optimization Tips
- Simplify System Prompt: Remove unnecessary descriptions, keep core instructions
- Limit Output Length: Use max_tokens parameter to control
- Clean Up History Messages: Keep only the last N turns, summarize the rest
- Monitor Usage: Regularly check token consumption reports on the DeepSeek platform
Cost Estimation
For example, with 1000 token input + 500 token output: cache miss ~¥0.002; cache hit ~¥0.00102. With 10,000 calls per day, each with 3000 tokens output, about ¥60/day (Flash with cache hit).