DeepSeek V3 vs V4 Flash vs R1: Which One Should Your App Actually Use?

Picking a model is easy when you have one choice. It gets hard when you have three — and they all claim to be the best.

Here's the decision framework I actually use after running DeepSeek V3, V4 Flash, and R1 side by side in production for three months.

The quick answer

• V4 Flash — your default for 90% of traffic. Fast, $0.14/M input, handles chat, extraction, classification, JSON output.
• R1 (reasoning) — only for tasks that genuinely need multi-step reasoning: code generation, math, complex agent planning. Budget 3-5x the latency and price.
• V3 — the legacy workhorse. Keep it as a fallback in your routing chain, not a primary.

How I decide (the 2-question test)

Question 1: Does the task need to "think" before answering?

If a human would answer in under 2 seconds, use Flash. If they'd grab a pen and paper, use R1. Most "hard" tasks are actually just long — they don't need reasoning, they need more context.

Question 2: What's the cost of being wrong?

Classification errors cost nothing. Code that doesn't compile costs hours. The more expensive the error, the more willing you should be to pay for R1.

The routing pattern that works

One gateway endpoint, three upstreams, and a simple rule:

default → V4 Flash (fast path)
math/code/reasoning keywords → R1
R1 timeout/failure → V3 fallback → Flash fallback

This cut our average latency 40% while keeping the "smart" answers for the tasks that need them. The key insight: you don't pick one model, you pick a routing policy.

What this costs in practice

For a product doing ~2M tokens/day: running everything on R1 would be ~$180/month. The routing policy above: ~$55/month. Same answers for the tasks that matter, ~70% cheaper.

If you want to test all three without the setup

I use ModelHub — one OpenAI-compatible key that routes to DeepSeek V3, V4 Flash, R1, and 40+ Chinese models. $5 free credit, no Chinese phone number needed for signup. I wrote a longer piece on the full architecture (rate limiting, error normalization, fallback chains) if you want the details.

---

What's your routing policy? Do you default to a reasoning model for everything, or gate it? Curious what others do in the comments.