One card in this batch moved 327,466,528,164 tokens in seven days [OURS]. The same card earned 129 search impressions in 28 days, with no clicks, at an average position of 15.5 [OURS: Search Console, 2026-09-24 to 2026-10-21]. It is the deepseek v4 pro 0813 pricing page, and the distance between those two numbers is the point. The card is served by OrcaRouter.
The traffic number
Three hundred twenty-seven billion tokens a week is not a rounding error in any catalogue. It is roughly thirty times the combined weekly volume of the GLM cards in this batch, and about fifty-seven times the OrcaCyber card’s 5.77 billion.
A single number that large on a card with four FAQs and no prose sections says something specific: the model is being reached through the API, not through search. Nobody arrives at 327 billion tokens by reading a marketing page.
That is normal for a flagship. It also means the card’s search performance and the model’s actual use are almost unrelated, which is why a page at position 15.5 can still be one of the busiest models on the site.
What the rate row charges
| Line | Rate per million |
| Input | $0.66 |
| Output | $1.98 |
| Cached read | $0.022 |
An output-to-input ratio of 3× and a cache read at 3.3 percent of input — both on the cheap side. The cache discount is 30×, which is steeper than the 10× that most cards in this catalogue offer.
At the volume this card actually carries, the difference between 10× and 30× is not academic. For 327 billion tokens a week, if a tenth of it were cache reads at the 10× discount, the weekly bill on those reads would be about $2,160; at the 3× that this card charges, it is about $720 [OURS arithmetic on the published rates].
The quiet caveat is the timed pricing: two windows a day double every line, which is the subject of the companion piece on this card.
Why the page is at 15.5
The query set explains it. Of 31 queries, the largest is the dated identifier:
| Query | Impressions | Position |
| deepseek v4 pro 0813 pricing | 54 | 20.7 |
| deepseek v4 | 15 | 10.3 |
| deepseek v4 pro 0813 playground | 9 | 20.8 |
| deepseek v4 pro api | 6 | 18.7 |
| deepseek v4 pro price | 6 | 11.7 |
| deepseek v4 pro 0813 price | 5 | 4.8 |
| deepseek v4 pro 0813 api | 4 | 27.0 |
The best-placed query in that list — `deepseek v4 pro 0813 price`, position 4.8 — carries 5 impressions. The query carrying 54 impressions sits at position 20.7, which is page three.
That is the shape of a page ranking on a snapshot identifier rather than on a model name. `deepseek v4` alone is 15 impressions at position 10.3, and it lands here because the page matches the string, not because it is the page Google considers the answer. The canonical page for the family is the undated `deepseek-v4-pro` card, and it carries 68 external links to this card’s 4 [OURS].
So the dated snapshot is doing what dated snapshots do: it accrues impressions for people who typed the date, at positions too deep to earn anything.
What a snapshot page is for
The card exists so that a team needing reproducible behaviour can pin a checkpoint. That is a legitimate product decision, and it makes the search situation a feature rather than a bug: the searcher who types a date is a searcher who wants that date, not the current model.
But it changes what the page should optimise for. A snapshot page does not need to outrank the canonical card; it needs to be findable by the people who know the date exists, and to state clearly what differs from the undated route.
The card states that in one line: it is the August 12 official release, superseding the April preview, serving a 1M-token context with up to 384,000 tokens of output [OURS]. What it does not state is whether the rates differ from the undated card — they do not, both list $0.66 and $1.98.
The takeaway
The DeepSeek V4 Pro 0813 card carries 327.5 billion tokens of seven-day traffic and 129 search impressions at position 15.5 [OURS]. It prices at $0.66 input, $1.98 output and $0.022 cached per million, with a 30× cache discount and two daily windows that double every line.
Traffic this size does not come from search. The page’s job is to be the reproducible pin for teams that need one, and its ranking on the dated query is doing exactly that — at a depth where it costs nothing and earns nothing.
Sourcing note: the three rates, the context and output limits, the FAQ and section counts, the fourteen benchmark rows, the seven-day token total and the timed-pricing flag are read from the OrcaRouter public model card for `deepseek/deepseek-v4-pro-0813` on 2026-10-23 [OURS]. Query, impression, click and position figures are Search Console for the 28 days to 2026-10-21 [OURS]. The comparison to other cards’ traffic, the cache-read arithmetic and the external-link counts are our own, computed from the same card file and the historical link ledger. No competing gateway or inference marketplace is named. All values checked 2026-10-23. See more