Which DeepSeek model?
Three questions. Most answers land on Flash.
Q1. Does your input include images?
Yes → Flash. V4 Pro does not accept images. Stop here.
No → continue.
Q2. What does a wrong answer cost you?
Nothing much (labels, formatting, extraction) → Flash.
Real money (code review, contracts, analysis, user-facing decisions) → V4 Pro.
Q3. Are you running high volume under tight latency?
Yes → Flash for the bulk path, V4 Pro only for the hard cases. Flash allows 2500 concurrent requests versus 500 for V4 Pro.
The two-model routing pattern
1. Cheap classifier (Flash) decides: simple or hard? 2. Simple → Flash ~80% of volume 3. Hard → V4 Pro ~20% of volume 4. Cache the system prompt for both
This typically lands at 25–35% of the cost of routing everything to V4 Pro, with most of the quality retained.
What you are choosing between
| Flash | V4 Pro | |
|---|---|---|
| Input (off-peak) | $0.15 | $0.66 |
| Output (off-peak) | $0.60 | $1.98 |
| Vision | Yes | No |
| Concurrency | 2500 | 500 |
| Best at | Pattern work, vision, volume | Reasoning, code, judgement |