Same Family, Two Very Different Jobs
DeepSeek shipped V4 as two models on April 24, 2026, and picking between them is mostly about one question: does your task actually need the extra reasoning? Both are 1M-context MoE models with the same OpenAI-compatible API — the difference is depth vs. price.
The Numbers
| V4 Flash | V4 Pro | |
|---|---|---|
| Params (MoE) | 284B / 13B active | 1.6T / 49B active |
| Input (per 1M) | $0.14 | ~$0.44 (promo) / ~$1.74 std |
| Output (per 1M) | $0.28 | ~$0.87 (promo) / ~$3.48 std |
| Cached input | ~$0.0028 (98% off) | ~$0.004 |
| Context | 1M | 1M |
| Max output | 384K | 384K |
At promo rates Pro is ~3x Flash; at standard rates it's ~12x on input. Flash is one of the cheapest quality models anywhere.
When Flash Is Enough (Most of the Time)
V4 Flash is the default. It handles classification, extraction, summarization, chat, RAG, coding subtasks, and cache-heavy repository work with frontier-adjacent quality. DeepSeek's own positioning says Flash matches Pro on simpler agent tasks at ~12x less cost. For high-volume and output-heavy workloads, Flash's $0.28 output is the killer number.
When Pro Earns It
V4 Pro is the escalation lane for genuinely hard work: competition-grade coding (LiveCodeBench 93.5, Codeforces ~3206), deep multi-step reasoning, and long-horizon agent tasks where a wrong answer costs more than the extra tokens. Reach for it only when Flash's quality demonstrably falls short on your evaluation — not by default.
The Right Setup: Flash Default, Pro Escalation
def deepseek_model(task):
return "deepseek-v4-pro" if task.needs_deep_reasoning else "deepseek-v4-flash"Start everything on Flash, promote only the tasks that fail your quality bar to Pro. This is the same tiered-routing logic that controls cost across any LLM stack — see best coding LLM API and how to reduce AI API costs.
Two Things to Know
- Thinking mode on either model bills at the same rate but burns 3–5x more tokens — off by default.
- Data residency: DeepSeek's official API stores data in China (a GDPR/compliance consideration). Self-host the open weights or use a zero-retention gateway. On LinkModel both models run behind one key with zero data retention by default, up to 30% below official. Also: legacy
deepseek-chat/deepseek-reasoneraliases retire July 24, 2026.
Bottom Line
- Flash for ~90% of work — cheap, fast, 1M context, great cache economics.
- Pro for the hard reasoning and competition-grade coding that justifies 3–12x the cost.
- Both, routed, is the cheapest way to keep quality where it matters.
Compare against the field in cheapest LLM API and DeepSeek V4 Flash vs GPT-4o.
Start free with a $1 credit and benchmark both on your workload.