Background
Effective August 17, 2026, DeepSeek V4 API pricing switched to a peak/off-peak time-of-use model, a structure familiar from electricity tariffs but new to LLM APIs. Peak hours are defined as Beijing time 09:00-12:00 and 14:00-18:00; all other hours are off-peak.
This is DeepSeek's fourth pricing adjustment of 2026.
Price Tables
V4 Pro (per million tokens, ¥)
| Item | Peak | Off-peak | Change at peak | |---|---|---|---| | Cached input | 0.30 | 0.15 | +1100% (from ¥0.025) | | Uncached input | 9.00 | 4.50 | +200% | | Output | 27.00 | 13.50 | +350% |
V4 Flash (per million tokens, ¥)
| Item | Peak | Off-peak | Change at peak | |---|---|---|---| | Cached input | 0.10 | 0.05 | +400% | | Uncached input | 3.00 | 1.50 | +200% | | Output | 9.00 | 4.50 | +350% |
Why Now — Demand Pressure
Two reference points explain the move:
- August 1: V4-Flash processed roughly 8 trillion tokens in a single day.
- August 4: the API went temporarily unavailable due to traffic.
- Off-peak discount gives an economic reason to migrate batch workloads to nighttime.
- Peak surcharge gives a compute reason: scarce compute capacity should be paid for by high-load customers, freeing low-latency budget for critical traffic.
- Cache-heavy work is pushed toward off-peak.
- Generation-heavy work pays the peak premium.
- Anthropic, OpenAI, Google Gemini: None currently use time-of-use pricing. Their variants are flat hikes (e.g., GPT-5.6 cached input at ~$0.05 per million tokens, roughly 1.25× DeepSeek's post-hike $0.04) or subscription/tier packages (e.g., Claude Max at $200 / $400).
- DeepSeek's approach mirrors the electricity tariff pattern familiar to Chinese consumers and the Spot Instance model used by Alibaba Cloud and Tencent Cloud for inference compute—but applied at the token level for the first time.
- The subsidy-for-scale era is ending; V4 enters scale-for-margin mode. Revenue elasticity turns sharply upward if daily volume keeps growing past 8 trillion tokens.
- Chinese LLM competition shifts from price cutting to pricing-mechanism design: time-of-use, tiered, subscription, token packs, and enterprise contracts will all run in parallel.
- Within ~3 months: whether more than two domestic vendors (Qwen / Zhipu / Kimi / Doubao) follow DeepSeek into time-of-use pricing—if yes, China's LLM API market officially enters an "electricity tariff" era.
- Within ~12 months: V4's actual peak-hour gross margin and whether off-peak pricing keeps developer goodwill intact.
The pricing logic follows:
The Real Blow Falls on Cache-Heavy Workloads
The 1100% headline understates the impact. Because cache-heavy agents, multi-turn dialogue, document review, and code review depend on prefix reuse, the *total* cost rise far exceeds the per-token percentage. DeepSeek is effectively redrawing a market line between cache-heavy and generation-heavy workloads:
The Four-Step Rollout
1. August 6 — API docs added a notice: *"Plans to raise API service pricing overall; expect a large increase, formal plan to follow."*
2. August 12 — Silent release of V4-Pro-0813: cached input raised from ¥0.025 to ¥0.10 per million tokens.
3. August 13 (evening) — Formal time-of-use announcement.
4. August 17, 00:00 — Effective.
The cadence was tighter and more telegraphed than any recent OpenAI pricing change.
Competitive Context
What Developers Should Do
1. Cache-heavy workloads: rebuild billing models; move everything possible to batch/offline. 2. Time-shift workloads: send nightly jobs, batch jobs, and deferrable jobs to off-peak for a flat 50% discount. 3. Multi-model fallback: during peak hours, route overflow traffic to domestic alternatives such as Qwen3.8-27B or Zhipu GLM-5.3.