The rate row on the DeepSeek V4 Pro 0813 card prints $0.66 input, $1.98 output and $0.022 cached per million tokens [OURS]. Those are the rates for seventeen hours of the day. For the other seven, every line doubles, and neither the rendered row nor any of the card’s four FAQs mentions it. The deepseek v4 pro price query lands on that row. The card is served by OrcaRouter.
Where the multiplier lives
The doubling is carried in the card’s structured price data as `timed_pricing`, timezone UTC, with two windows:
| Window (UTC) | Hours | Multiplier |
| 01:00 – 04:00 | 3 | 2× |
| 06:00 – 10:00 | 4 | 2× |
| All other hours | 17 | 1× |
Seven hours out of twenty-four is 29.2 percent of the day. During those hours input is $1.32, output is $3.96 and the cached read is $0.044.
The card is not hiding this deliberately. A rate table has one column per rate, and a second rate for the same token would mean duplicating every row with a time qualifier. The multiplier lives one level down, where a routing layer can read it and a human reading the page cannot.
What it does to a flat workload
Take requests spread evenly across the day.
| Table rate | Real average | |
| Input | $0.66 | $0.8525 |
| Output | $1.98 | $2.5575 |
| Cached read | $0.022 | $0.0284 |
The table’s number is 77.4 percent of the real one on all three lines, because the multiplier scales everything equally [OURS arithmetic on the published rates].
A workload that budgets from the printed row understates its bill by 29 percent. That is the same magnitude as a mid-tier caching decision, and it is invisible on the page.
What it does to a scheduled one
The more useful framing is the batch job, because the windows are known in advance.
Take 500,000 input tokens and 100,000 output tokens per run, once a night:
| Schedule | Input | Output | Per run |
| 22:00 UTC | $0.33 | $0.198 | $0.528 |
| 02:00 UTC | $0.66 | $0.396 | $1.056 |
Thirty runs a month: $15.84 against $31.68 [OURS arithmetic on the published rates]. The job did not change; the clock did, and the bill doubled.
That is a scheduling decision worth making explicitly for any batch workload, and worth making once rather than per-job. The windows are fixed, so a queue that drains overnight can simply be told to avoid 01:00–04:00 and 06:00–10:00.
Where the multiplier is a family trait
In a 203-card snapshot of this catalogue taken on 2026-10-21, nine of the eleven DeepSeek rows carried `timed_pricing`, and no card from any other provider did [OURS].
That asymmetry has a practical consequence for failover. Routing layers compare rates to decide which route to prefer. Put a DeepSeek card beside a second provider and only one side of the pair varies by clock — so at 02:00 the failover looks relatively cheaper than it is, and at 22:00 the DeepSeek primary looks cheaper than the average it will actually bill.
Neither comparison is wrong at the instant it is made. Both are wrong as a basis for a monthly budget.
Reading a card that prices by the hour
Three rules.
Check the structured price, not the rendered row. The multiplier is in the data behind the page. A reader who works only from the visible rate row never sees it.
Compute the blended rate before you compare cards. $0.66 against another model’s $0.80 is not a comparison if one of them averages $0.8525 across the day.
Schedule against the windows, not around them. Two fixed bands are easy to avoid for batch work and impossible to avoid for interactive traffic. The distinction decides whether the multiplier costs you anything at all.
The takeaway
DeepSeek V4 Pro 0813 prints $0.66 input, $1.98 output and $0.022 cached per million tokens, and doubles all three for seven hours a day across two UTC windows [OURS].
The blended cost of a flat workload is 29 percent above the printed rate, and a nightly job moved two hours can double in price. The multiplier is in the pricing data, not on the page — which is exactly where a routing layer can read it and a budget spreadsheet cannot.
Sourcing note: the three base rates, the `timed_pricing` windows and multiplier, the context and output limits, and the fourteen benchmark rows are read from the OrcaRouter public model card for `deepseek/deepseek-v4-pro-0813` on 2026-10-23 [OURS]; the rates are the provider’s published prices passed through with $0 markup [VENDOR]. The count of cards carrying timed pricing comes from a 203-card snapshot taken on 2026-10-21. The blended-rate arithmetic, the per-run cost table and the failover argument are our own. The impression and position figures are Search Console for the 28 days to 2026-10-21 [OURS]. No competing gateway or inference marketplace is named. All values checked 2026-10-23.