The same token now has business hours. On August 13, DeepSeek put V4-Pro-0813 into general availability with a peak/off-peak price structure: off-peak at 0.66 per million input tokens and 1.98 per million output, with both rates doubling during peak windows - 1.32 and 3.96. Peaks run 01:00-04:00 and 06:00-10:00 UTC, and weekend off-peak pricing began August 23. If you buy tokens in bulk against a meter that does not record the hour, nothing changed. If you run a router, a new dimension just entered the routing decision: the clock.
How time-of-use rates actually work
Electricity utilities have run the same mechanism for decades: time-of-use (TOU) pricing. The grid's marginal cost varies by hour - midday demand, when everything draws at once, is served by expensive peaker plants, while off-peak demand rides cheap baseload. The utility passes that through to the price tag and effectively charges for inflexibility: if you can shift your load, you get the cheap price; if you cannot, you pay peak.
Two conditions make TOU work:
- Capacity is fixed and sunk. The generator either runs or idles. A megawatt at 3 a.m. has a marginal cost near zero; the same megawatt at noon carries the full cost of the peaker.
- Some load is shiftable. If no customer can shift, TOU does not work - the utility just raises the flat rate instead.
Map that onto a data center running inference:
- GPUs are fixed and sunk. A card that produced nothing this hour paid for itself this morning; the marginal cost of serving another token is memory and power, not silicon.
- Inference demand is genuinely variable by hour. Interactive traffic concentrates in the buyers' local daytime. Batch, agents, and scheduled jobs can move.
DeepSeek's TOU structure is an implicit admission that it has hours where the fleet is under-utilized - idling while fully paid - and hours where demand strains comfortable utilization. The 2x peak spread is the utility's tell: at the peak window, the marginal machine is expensive, and the price says so.
Notice the shape of the windows. 01:00-04:00 UTC is late evening across the Americas; 06:00-10:00 UTC is European morning. The peaks are organized by geography, not by algorithm. Euro-American interactive demand is the thing being priced.
Why nobody else has done it yet
OpenAI, Anthropic, and Google ship flat rates as of this writing. Two reasons matter, both of which could break within a year:
- Their demand shape is flatter. Multi-model APIs with global users average out the hour effect. TOU only pays where the provider can see its own peaks, and a single-model retail product sees them loud and clear.
- Flat pricing is a simpler selling story. TOU exchanges a simple line - "per million tokens" - for revenue optimization. A provider with an API-first business can absorb that trade; a consumer brand cannot without a service-desk incident.
There is a third reason that matters most to you, the buyer: the moment a second provider adds hour-dependent pricing, "which model, at what price, at what time" becomes a three-variable decision - model, rate, hour - and a static routing table stops being sufficient.
What changes inside the routing decision
The routing decision was: pick a model that meets the quality bar at the best rate. For shiftable work, it is now: pick a model that meets the quality bar, at the cheapest price point within the window that meets the deadline. Latency-sensitive work (user-facing chat) happens when users are awake - that is precisely peak - so TOU is a tax on interactive traffic and a real recurring discount on scheduled work: nightly digests, data-pipeline enrichment, agent warmup, index building.
One compounding effect is worth a line: DeepSeek reports cache hits cutting input cost by as much as roughly 97 percent. Off-peak plus cache-hit is the strongest position in the whole matrix, and it is the position a router can actually sit in.
What to do
- Classify your traffic as shiftable or immovable. Anything with a deadline and a slack window gets an off-peak preference where the provider supports it.
- Anchor peak windows in UTC in your scheduler - not local time - and give jobs a 30-60 minute buffer at window boundaries.
- Keep the hour as a first-class routing signal that can be enabled per task class. Mappace exposes effective per-hour rates per model so a scheduler can optimize at runtime instead of at deploy time.
- Watch the other major providers. A second adopter is the moment the industry-wide inference grid gets deep enough for demand pricing to become the default and flat rates the exception.