The 75% cache discount on R1 inputs ($0.14 vs $0.55) is the biggest cost lever for R1 workloads. For reasoning tasks that reuse a stable system prompt across many requests, cache hits can dominate the input cost. R1 is most cost-effective when the reasoning work is genuinely hard — math, code, multi-step debugging. For routine chat or RAG, V3.2 or V4 Flash is the better economic choice (TokenMix, April 2026).
The V3, V3.2, V4 family tree
DeepSeek offers five live models in the family as of September 2026, with overlapping price-performance that can confuse first-time buyers.
V3 at $0.27 input / $1.10 output is the legacy general-purpose model with a 128K context window. The DeepSeek API now routes chat workloads to V4 Flash by default; V3 is maintained for compatibility but no longer the recommended starting point. V3.2 at $0.28 input / $0.42 output has the same input price as V3 but 2.6x cheaper output, with improved coding and reasoning performance and a 32K max output (vs V3's 16K) (DeepSeek API docs, September 2026).
V4 Flash at $0.14/$0.28 is the new default chat model. V4 Pro at $0.44/$0.87 is the flagship 1M-context model. R1 at $0.55/$2.19 is the reasoning specialist. R1-0528 at $0.50/$2.15 is a May 2025 snapshot of R1 with improved reasoning performance. For most production workloads, the recommended path is V4 Flash for general chat, V3.2 for coding, V4 Pro for enterprise flagship, and R1 for reasoning-heavy workloads (aicostcheck.com, April 2026).
Off-peak discounts and the pricing model
DeepSeek runs a peak/off-peak pricing model that cuts the bill by 50% during off-peak hours.
DeepSeek's pricing model charges 50% less during off-peak hours for the same workload. Off-peak hours are typically 16:30-00:30 UTC (00:30-08:30 Beijing time). The model is opt-in: developers can choose to route traffic during off-peak hours by setting the appropriate parameters. For batch processing, embeddings generation, and other non-real-time workloads, off-peak routing cuts the bill by 50% on top of any prompt-cache savings (DeepSeek API docs, September 2026).
The 50% off-peak discount is the same magnitude as OpenAI Batch API (50% off standard) and Google Vertex AI Batch Mode (50% off standard). DeepSeek's pricing model does not include a separate Batch API — off-peak routing is the equivalent mechanism. For DeepSeek V4 Flash, the off-peak rate is $0.07 input / $0.14 output per million tokens — cheaper than any other provider's standard rate (DeepSeek API docs, September 2026).
What enterprise buyers should do next
Three actions for organisations evaluating DeepSeek API pricing in 2026.
- Pick V4 Flash as the default. For most production workloads, V4 Flash is the right starting point. Move to V3.2 when the task needs better coding or longer output (32K). Move to R1 when the task needs deeper reasoning and the cost of CoT tokens is justified by accuracy gains. Move to V4 Pro when you need 1M context and enterprise SLAs.
- Use prompt caching aggressively. DeepSeek's cache rate of $0.0028 per million tokens on V4 Flash is the cheapest in the LLM market. A 50K-token system prompt reused 100x per hour costs $0.014 per hour on V4 Flash versus $0.50 per hour on V3 without caching.
- Plan for the China data-residency question. DeepSeek is hosted in China. For US-regulated workloads (HIPAA, FedRAMP, CMMC), use DeepSeek for development and cost-optimised non-sensitive workloads, but route sensitive production data through Azure, Bedrock, or Anthropic direct. For non-regulated workloads, DeepSeek's 14-71x cost advantage is real and durable.
What to watch next
Three near-term datapoints. First, DeepSeek's announced V4 Pro API service after May 31, 2026 — the 75% promotional discount ended May 31, 2026, and standard pricing applies since then. Second, the introduction of DeepSeek on AWS Bedrock and Azure AI Foundry as a hosted model — currently DeepSeek is direct-only. Third, the release of DeepSeek R2 — DeepSeek typically ships a successor reasoning model every 8-12 months; R2 is plausible for Q4 2026.