Following DeepSeek's latest pricing update, a wave of panic swept through tech communities: headlines screamed about "+1,100%" increases and 12.1x price jumps. Executives and CTOs immediately scrambled to recalculate budgets, bracing for massively inflated inference bills. However, if you take a realistic enterprise workload profile and do the math carefully, the actual cost multiplier lands between 2.3x and 2.9x—nowhere near an elevenfold surge.
The DeepSeek team transitioned pricing from a flat rate to a two-tier model on August 16 at 16:00 UTC. Peak hours are now exactly twice as expensive as off-peak hours, with no intermediate tiers. The sensational 12.1x leap exists strictly in a single line item out of six: cached input for the Pro model, where the price nominally rose from $0.003625 to $0.044 per million tokens.
Anatomy of a single price cell
The $0.003625 figure had been in place since April 26, when the provider temporarily slashed cache hit prices tenfold. Before that promotional discount, cached input on Pro cost $0.03625, meaning the current peak rate represents just a 21% increase over the original spring baseline. On the Flash tier, cache is still half the price of its launch rate, so the alarming headlines about a twelvefold hike merely reflect the expiration of a temporary discount on the cheapest line item.
In an actual corporate bill, this line item carries negligible weight because cached input was microscopically cheap to begin with.
The share of cached input in a chatbot assistant's bill rose from 0.59% to 2.04%, and in RAG scenarios it increased from 2.79% to 9.71%, making a total cost surge of 12x mathematically impossible.
The other five price items went up by 3.0–5.0x during peak hours and 1.5–2.4x off-peak—and these represent the bulk of actual inference spending. To see a 12x bill increase in practice, incoming requests would have to consist almost entirely of cache hits paired with an output length of exactly one token. In real-world enterprise production, such synthetic scenarios simply do not exist.
Beijing lunch breaks and global time zones
The second part of the operational equation lies in the geography of peak windows, documented as 01:00–04:00 and 06:00–10:00 UTC. The two separate windows look strange only until you convert them to Beijing time (UTC+8). In the local schedule, they align with standard 09:00–12:00 and 14:00–18:00 office hours, with the two-hour gap matching a typical Chinese corporate lunch break.
DeepSeek configured its pricing grid around domestic traffic, leaving the rest of the world to deal with the schedule without local adjustments. For European and nearby time zones, this creates advantageous scheduling opportunities: off-peak rates cover substantial portions of the working day and night.
Shifting batch computations and background jobs in your task scheduler by just a few minutes to align with off-peak windows brings expenses down to baseline rates. Instead of a catastrophic 1,100% surge, businesses face a manageable 2.3–2.9x budget increase—provided the technical leadership understands how to configure queues around Beijing's lunch schedule.