DeepSeek V4 Pro Pricing: Peak & Off-Peak Rates (Aug 2026)

DeepSeek introduced peak and off-peak rates for the whole V4 lineup, effective 16:00 UTC on August 16, 2026. V4 Pro's per-1M-token prices are now tiered by hour^[1][2].

Off-peak is exactly half the peak price, so the scheduling decision is worth roughly 2x on every token. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; everything outside those windows is off-peak[7].

For comparison, V4 Flash runs at $0.22 (off-peak) / $0.44 (peak) per 1M input on a cache miss and $0.66 / $1.32 per 1M output — about one-third of Pro on most lines.

Before August 16, V4 Pro had a single flat rate: $0.435 input (cache miss), $0.003625 input (cache hit), and $0.87 output. The new schedule raises those numbers and splits them by hour^[2].

Every line went up — there is no tier where the new price is lower than the old flat rate. Output is the biggest absolute jump ($0.87 → $3.96 peak), while cache-hit input is the biggest percentage jump^[2].

If you are comparing V4 Pro against closed flagship models or planning a Flash-vs-Pro routing strategy, use the off-peak numbers for baseline capacity and the peak numbers for worst-case budgeting.

A widely shared headline claimed DeepSeek raised V4 Pro prices by 1,100%. The real number depends on which line you are looking at — and that specific figure is real but only for the most extreme case^[2][7].

The +1,114% figure is the cache-hit input price at peak hours: $0.003625 → $0.044 per 1M tokens. It is technically accurate and utterly misleading as a headline, because cache-hit input is the cheapest line in the entire price table — even after a 12x increase it is still far cheaper than a cache miss.

The lines that actually affect most production budgets are output (+128% off-peak, +355% peak) and input on a cache miss (+52% / +203%). For heavy cache users, the new cache-hit rates still make prefix caching the dominant cost-saving lever — just a smaller one than before.

• Output: +127.6% off-peak / +355% peak
• Input cache miss: +51.7% off-peak / +203% peak
• Input cache hit: +507% off-peak / +1,114% peak
• Headline '1,100% hike' = cache-hit input at peak, the cheapest line in the table

Even after the increase, cache-hit input is 30x cheaper than a cache miss at the same hour ($0.022 vs $0.66 off-peak). If your workload reuses prefixes — multi-turn agents, RAG, code review loops — caching dominates the bill^[2].

• Prefix caching is automatic — no headers or SDK changes
• The hit price applies to any replay of the same prefix, including partial matches
• The context caching guide covers how to structure prompts for maximum hits
• Off-peak cache hits ($0.022) are the single cheapest way to run V4 Pro

# Cost per 1M input tokens, off-peak
cache_hit  = 0.022   # same prefix replayed
cache_miss = 0.66    # first-time prefix

# A 20-turn agent loop with a 50K-token shared prefix:
#  turn 1 costs miss price for the prefix, turns 2-20 cost hit price
saved = 19 * (cache_miss - cache_hit)   # ≈ $12.1 per 1M prefix tokens

Concrete planning numbers: V4 Pro allows 500 concurrent requests (Flash allows 2,500), and pricing is now hourly, so capacity planning should account for when your traffic runs^[2].

A practical budget model for a typical agentic workload: assume 40% cache-hit input, 60% cache-miss input, and output at 1.5x input volume (thinking tokens count as output). At off-peak hours that lands around $1.45 per 1M input-equivalent tokens; at peak it is roughly double.

If your workload can defer, the peak vs off-peak scheduling guide shows how to shift batch jobs to the cheap hours. For always-on interactive traffic, the cache-hit rate is your main lever — structure conversations to keep a stable system prompt and tool schema so prefixes replay.