DeepSeek's 50% off-peak discount is a price increase, and scheduling is now a cost control

DeepSeek has split every token price into peak and off-peak tiers, with off-peak rates 50% below peak and the change taking effect at 16:00 UTC on August 16, 2026 ^[1]. According to a dev.to writeup by jamilxt, who had been running V4 Pro through a Spring AI app and knew his old rates, every new off-peak number still sits above the old flat rate, which makes this a repricing with a discount bolted on rather than a discount ^[8][11].

The mechanics first. Peak hours are 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, and everything else is off-peak, which the pricing page counts as 17 of 24 hours ^[2][3]. That is seven peak hours, or about 29% of the clock ^[1], with an odd two-hour off-peak island from 04:00 to 06:00 UTC sitting between the two peak blocks ^[2]. One commenter in the Hacker News thread noted that the peak windows line up with China's workday and fall at night in the US and Europe ^[17], which is the shape you would expect from pricing driven by domestic demand.

The rates: V4 Flash off-peak is $0.22 per million input tokens on a cache miss, $0.66 output, $0.007 cache hit, and peak is exactly double at $0.44, $1.32 and $0.014 ^[4][5]. V4 Pro off-peak is $0.66, $1.98 and $0.022, with peak at $1.32, $3.96 and $0.044 ^[6][7]. Against the old flat rates of $0.435 input and $0.87 output for Pro ^[8], off-peak is roughly 1.5x and 2.3x, and peak is roughly 3x and 4.6x ^[11]. Flash, previously $0.14 and $0.28 ^[10], is about 1.6x off-peak and 3x to 4.7x at peak ^[12]. Because the best available first-party rate is now more than double the old output price, no amount of scheduling recovers the old bill ^[6].

The item that actually changes system design is cache hits. Pro cache-hit pricing went from $0.003625 to $0.022 off-peak and $0.044 at peak, roughly 6x and 12x ^[9][13]. The sharpest comment in the thread, per the writeup, points out that cache hits can be 90% or more of input on long agent sessions ^[15], and that this "will put an end to those 'I used 1 billion tokens and spent $4' reports" ^[16]. Run the blend on a session that is 90% cache hits: input cost per million goes from about $0.047 under the old flat rate to about $0.086 off-peak and about $0.172 at peak, so 1.8x and 3.7x ^[3]. One developer estimated a $5 monthly bill becoming $15, a 3x jump ^[19][5].

Comparison shopping is now time-dependent. Peak Flash output at $1.32 is above the cheapest OpenRouter provider at $0.16, roughly 8x ^[20][4], and because the weights are open, other providers were still selling the same models at the old prices or lower when the thread was running ^[18]. The change is first-party only.

Worth noting that the writeup is a reading of the price table plus two days of Hacker News math on a 128-point, 183-comment thread; the author had not put a real invoice through the new billing when he published ^[14][21]. What to watch: whether the peak windows move once DeepSeek sees how much load shifts into the 17 cheap hours, whether third-party hosts of the same open weights hold their old rates, and whether your own traffic mix is dominated by cache hits, because that is where the multiplier is worst.