DeepSeek V4 API Switches to Peak Pricing, With Pro Cache-Hit Rates Rising to as Much as 12.1× Their Previous Level

DeepSeek activated its revised V4 API pricing at 16:00 UTC on August 16, replacing its previous flat, around-the-clock rates with peak and off-peak tiers. According to the official documentation, peak periods run daily from 01:00–04:00 and 06:00–10:00 UTC; prices are halved during all other hours. V4 Pro output tokens now cost $1.98 per million off-peak and $3.96 at peak, while cache-miss input costs $0.66/$1.32. The corresponding V4 Flash prices are $0.66/$1.32 for output and $0.22/$0.44 for cache-miss input.

From an engineering perspective, cache hits deserve even more attention. The previous Pro rate was $0.003625 per million tokens. The new off-peak rate is $0.022, about 6.1 times the previous price, while the peak rate is $0.044, about 12.1 times higher. Flash has likewise risen from $0.0028 to $0.007/$0.014. This changes the cost model for agents: coding agents typically resend large system prompts, tool definitions, library contents, and conversation history, so bills may be driven primarily by cache-read volume rather than new input or output.

One community user recalculated costs using 650 real requests and 155.9 million input tokens, reporting that a workload with a 98.09% cache-hit rate became roughly 2.7 times more expensive. This is only one example and cannot be generalized to every service, but it shows that comparing only headline input/output prices can underestimate agent costs. Teams should replay their billing calculations using actual usage fields, tracking hits, misses, output, and UTC time windows separately. Deferrable batch evaluations and data-processing jobs should also be considered for off-peak scheduling. Cache hits remain far cheaper than misses, so caching should not be disabled indiscriminately. The next questions are whether gateways can become aware of time-of-use pricing and whether model providers will adjust third-party resale rates.