On 16 August 2026 DeepSeek split its API pricing into Peak and Off-Peak windows. Peak runs 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday; outside it is Off-Peak at half the rate. The published price sheet lists the exact figures.
1. DeepSeek's New Pricing Structure
The company replaced its previous flat rate with prices that vary by 100% (2x) depending on the time of day, retiring the legacy names deepseek-chat and deepseek-reasoner in favour of the V4-Flash (low latency) and V4-Pro (deep reasoning) models.
For V4-Flash Off-Peak: Cache Hit $0.003, Input $0.15, Output $0.60 per 1M tokens. Peak doubles those figures. The break-even math uses the published price sheet, not an invented discount.
Prefix caching reuses a matched prefix at a fractional cost. The discount applies only when the exact prefix repeats; injecting the current timestamp or a session ID at the top of the prompt invalidates the cache.
3. ROI Analysis: DeepSeek vs. Western Models
While Claude Opus 5 costs $15/$75 per 1M tokens and GPT-4o is around $2.50/$10, DeepSeek V4-Flash operates at fractions of a cent. For customer service automation, WhatsApp CRMs and large-scale triage, the monthly savings enable unbeatable operating margins.
4. How to Save More than 95% with Prompt Caching
By keeping system prompts static, placing dynamic variables at the end of messages and ordering histories chronologically, cache hits reduce the input cost to just $0.007 per million tokens on V4-Flash.
5. Token Financial Management Strategies
Schedule asynchronous jobs (such as reports and data indexing) to run outside peak hours and configure intelligent proxies (such as LiteLLM) to trigger V4-Pro only when the task's complexity requires deep reasoning.
Reference: DeepSeek pricing. Reference: OpenAI prompt caching.