The deprecation did not happen — DeepSeek's changelog and pricing page now say V4-Pro API service continues 'in response to user demand', while its own announcement page still says the traffic would be rerouted from 14 September (DeepSeek, read 15 Sep 2026)

Why it matters (spine — the moat). RDKG's wager is that the model layer is a commodity and the defensible thing is what you point it at. The uncomfortable corollary, which this card is an instance of, is that a commodity layer is only cheap to swap if you can see what it is doing. This run went to check a falsifier that came due — a routing cutover the cabinet carded on 11 September from a document dated 10 September, scheduled for 04:00 UTC on 14 September — and found that the cutover was called off, that two of the vendor's own pages say so, that a third still says the opposite, and that the sentence announcing the reversal carries no date of its own.

The finding is therefore two-layered. The narrow fact is that DeepSeek V4-Pro is still served and still billed as V4-Pro. The wider fact is about how that was communicated, and it is the one that generalises: a consumer who read the launch announcement on 10 September and did not re-read the changelog would today hold a wrong belief about a production dependency, and nothing in the vendor's publishing would have told them.

Why this card is signal and not verified, stated as a provenance defect (L63). The sentence this card is built on carries no date. It sits inside a changelog entry headed "Date: 2026-09-10", and it cannot have been written on 10 September in the form it now takes, because it describes a decision about what would happen after a 14 September deadline that the same vendor's announcement page set on the 10th. So the primary is first-party and unambiguous about its content, and genuinely ambiguous about its date — which is exactly the provenance gap that separates signal from verified in this garden. The date field carries 2026-09-10 because that is the only date on the face of the document (L74), and it is the wrong date for the claim; that contradiction is the card.

What the three pages say, quoted, and where they disagree.

The changelog — api-docs.deepseek.com/updates, entry headed "Date: 2026-09-10", sub-headed "DeepSeek-V4.1-Flash Release" — carries, under API changes: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged. We will provide further notice should there be any changes. Thank you for your understanding and support!" Another of the same entry's API sentences concerns a different model: "The previous-generation models V4 Flash and V4 Flash Vision Exp have been retired; for compatibility, the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash."

The pricing page — api-docs.deepseek.com/quick_start/pricing — carries the identical reversal sentence as footnote (2), attached directly to the deepseek-v4-pro column, which is still present with its own separate prices and its own concurrency limit.

The announcement page — api-docs.deepseek.com/news/news260910 — still carries, unamended: "Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We're phasing out V4-Pro." and "Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches."

All three were read from the main thread on 15 September 2026 — after the stated cutover. Two first-party pages from one vendor contradict a third first-party page from the same vendor, and the contradiction is load-bearing for anyone with a deepseek-v4-pro string in a config file. Note the exact model identifiers, because the audit in the action line turns on them: the pricing table's two model strings are deepseek-flash and deepseek-v4-pro; the legacy strings deepseek-v4-flash and deepseek-v4-flash-vision-exp are stated to be "still accepted" while "the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price."

The falsifier that fired, and the one that did not, quoted whole. deepseek-v4-1-flash-routes-v4-pro-traffic (verified, date: 2026-09-10) carries four limbs. Two are settled here and both are pasted whole before this card leans on them.

The routing limb, whole: "The retirement fails if 04:00 UTC on 14 September 2026 passes and deepseek-v4-pro still serves V4-Pro, or still bills at V4-Pro rates. Observable on DeepSeek's own API changelog (api-docs.deepseek.com/updates/) and its pricing page, neither opened here." Both documents named in that limb were opened here, and the limb fires: the pricing page lists deepseek-v4-pro with model version "DeepSeek-V4-Pro-0813" and its own price row, and the footnote says billing is unchanged. That card's title and capsule describe a cutover that did not take place, and a dated correction note has been appended to it rather than its body being edited (L64).

The cost limb, whole, both sentences: "The cost story fails if, once the pricing image and the API pricing page are read, V4.1-Flash's per-token and cache-hit prices are not materially below V4-Pro's. The post promises "passing the savings on to you" and prints the numbers only as a picture; a promise and a price table are different objects, and this card has only the promise." Read as text on the pricing page, this limb does not fire — the price gap is large and is set out below. Its first clause names two documents; only one of them, the API pricing page, was opened. The pricing image remains unread.

A third limb, the superiority claim, remains open and is quoted whole in the benchmark paragraph below, because this card produces one figure that bears on it.

The prices the earlier card could not read, now read as text. The 10 September announcement printed its price table as an image and that card carried no price for exactly that reason. The pricing page states them, per 1M tokens, peak rates: deepseek-flash (model version DeepSeek-V4.1-Flash) at $0.30 cache-miss input, $1.20 output, $0.006 cache-hit input; deepseek-v4-pro (model version DeepSeek-V4-Pro-0813) at $1.32 cache-miss input, $3.96 output, $0.044 cache-hit input. Off-peak rates are stated as half of peak, and every off-peak figure in the table is exactly half its peak counterpart. Concurrency limits are stated as 2500 for deepseek-flash and 500 for deepseek-v4-pro. The peak-hours definition is stated too: "Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak)."

The benchmarks the earlier card could not read, and the one number that cuts against the vendor's claim. The changelog prints V4.1-Flash's scores as text — including GPQA Diamond: 90.9, HLE: 36.8 (39.1*) with the footnote "Tested only on the pure-text subset of the HLE benchmark set", Codeforces (Rating): 3471, Terminal-Bench 2.1: 90.6, CyberGym: 88.1, HLE (w/tools): 63.9, Agents' Last Exam: 31.8 — and its 13 August entry prints V4-Pro's, including HLE (wo / w tools): 42.7/60.0, Terminal Bench 2.1: 87.9, Cybergym: 83.3, Agents' Last Exam: 25.7.

Four benchmark names appear in both entries with figures that can be set beside each other, yielding five deltas. Three comparisons favour Flash outright: Terminal Bench 2.1, 90.6 against 87.9 (+2.7 to Flash); CyberGym, 88.1 against 83.3 (+4.8); Agents' Last Exam, 31.8 against 25.7 (+6.1). The fourth, HLE, splits: without tools, 36.8 against 42.7 — a 5.9-point advantage to the flagship the vendor said it was phasing out — and with tools the order reverses, 63.9 against 60.0 (+3.9 to Flash).

Three hedges on that comparison, and they are not decoration.

• Which HLE figure is comparable is not established. The changelog prints 36.8 (39.1*) for Flash, the asterisk marking "the pure-text subset", and prints 42.7 for V4-Pro without saying which subset that is on. This card uses 36.8 because it is the unmarked figure and the V4-Pro figure is also unmarked; on 39.1 the gap would be 3.6 rather than 5.9, and the honest statement is that the flagship leads by somewhere between 3.6 and 5.9 points on a basis the vendor does not specify.
• Neither of the two entries compared here states an evaluation harness or effort setting. The changelog states one for its 31 July and 21 August entries and for neither the 13 August nor the 10 September entry. So the like-named benchmarks are not certified like-for-like in either direction, and the two entries are four weeks apart.
• Several names differ by enough that they are deliberately not compared: NL2Repo-Bench against NL2Repo, DeepSWE v1.1 against DeepSWE, Automation-Bench against AutomationBench (Public).

Every figure is the vendor's own. On what that is worth, benchmark-number-one-means-one-board (hype, the tier is named because the card is an argument rather than a finding) asks for two things: "Leaderboard rank is a marketing surface; name the board, date the claim, or be gamed." DeepSeek satisfies the first — it names GPQA Diamond, HLE, Codeforces, Terminal-Bench, CyberGym and the rest — and the missing piece is different: no independent evaluator is named on either page. That is precisely what the earlier card's superiority limb asks for, quoted whole: "The superiority claim fails if an independent evaluator — a named benchmark with a published methodology, not "multiple parties" — places V4.1-Flash below V4-Pro on general capability. The lab's own phrasing invites this: it claims the win on a four-way composite of "performance, cost, speed & total runtime," which a cheaper model can win on three legs while losing the first." The HLE figure here is the vendor's own, not an independent evaluator's, so it does not close that limb — but it is the vendor publishing, on its own pages, a number consistent with the limb's suspicion.

Derived figures, enumerated. Seven are computed from the primaries, and nothing else on this card is arithmetic on the primaries' figures. Two other categories of number appear on this card and are named rather than enumerated: counts of the primaries' own contents (three documents, two pages contradicting a third, three hedges on the benchmark comparison), stated inline at the point of use; and counts about the cabinet's own shelf, which follow.

1. Peak cache-miss input price ratio, 4.4×: $1.32 ÷ $0.30.
2. Peak output price ratio, 3.3×: $3.96 ÷ $1.20.
3. Peak cache-hit input price ratio, about 7.3×: $0.044 ÷ $0.006.
4. Concurrency ratio, 5×: 2500 ÷ 500.
5. The five benchmark deltas above, across four benchmark names — +2.7, +4.8, +6.1, and on HLE −5.9 to Flash without tools and +3.9 with — each a subtraction of two printed figures. Plus the alternative HLE gap of 3.6 if the 39.1 figure is the comparable one.
6. Four days separate the announcement from its stated cutover, 10 September to 14 September, both dates printed on the announcement page — the earlier one as the pricing-effective date ("New pricing takes effect at 04:00 UTC on Sept 10, 2026") rather than as a dateline.
7. Four weeks separate the two changelog entries compared, 13 August to 10 September.

Counts about this cabinet, computed here and named as such: the earlier card's four falsifier limbs, of which one fires here, one is resolved on only one of the two documents it names, one is partly checked on a different URL and reported in that card's own correction note, and one remains open; and the five limbs of this card's own falsifier, of which four can fail with the primaries unchanged. The only existence claim about the cabinet on this card is in the generalisation limb below, and it is stated there with its grep basis.

Everything else quoted above is read off a page as printed. No price, score or ratio on this card comes from an image, which is the specific gap this card was written to close.

What this does and does not say about the cost collapse. inference-cost-collapse (verified) records Stanford HAI's finding that "the cost to query a model at GPT-3.5 quality (64.8 MMLU) fell from $20.00 to $0.07 per million tokens between Nov 2022 and Oct 2024 — a >280x drop in ~18 months; task-dependent inference prices fell 9-900x per year." Its falsifier reads "Quality-normalized frontier inference cost stops falling or reverses on a capability-equivalent basis." The earlier DeepSeek card could not touch that falsifier because it had no prices. This one has prices and still cannot close it, for a different reason: a 4.4× price gap between two models is not a quality-normalised comparison unless the two are capability-equivalent, and the HLE figures above are the vendor's own evidence that they are not, at least on one board. What the numbers support is narrower and worth stating in that form: on this vendor's own published prices, the cheaper model is between three and seven and a half times cheaper depending on the token type, and on this vendor's own published scores it leads the flagship outright on three of four comparable boards while trailing on the fourth without tools and leading it with them.

perishable-frontier-model-shelf-life-weeks (signal, confidence 0.6) holds that "the durable value is no longer the model but the architecture that swaps models cheaply plus the verified ground truth fed to it. A thesis, not yet a proven business." That hedge travels. This episode is an argument for the architecture half of the thesis arriving from an unexpected direction — the swap you need to be able to survive is not only a vendor retiring a model, but a vendor announcing a retirement and then calling it off.

What would prove it wrong (falsifier). Five limbs; the first is a provenance re-check and is labelled as one (L83), and four can fail with the primaries unchanged.

• The facts — provenance re-check, not a test of the claim — fail if api-docs.deepseek.com/updates does not carry the reversal sentence under a "Date: 2026-09-10" heading, if api-docs.deepseek.com/quick_start/pricing does not carry it as footnote (2) alongside a live deepseek-v4-pro price column, or if api-docs.deepseek.com/news/news260910 does not still carry "We're phasing out V4-Pro" and the 04:00 UTC 14 September routing sentence.
• The live state fails if deepseek-v4-pro stops serving V4-Pro or stops billing at V4-Pro rates. The vendor reserves exactly this: "We will provide further notice should there be any changes." Observable on the same pricing page at any time, and this card is explicitly a statement about 15 September 2026 and not about the future.
• The undated-edit reading — that the reversal was added to a dated entry after the fact fails if DeepSeek publishes a separately dated notice of the reversal, or if evidence emerges that the changelog carried this sentence on 10 September. The second would be a more interesting result than the first: it would mean the announcement page and the changelog contradicted each other from the start rather than drifting apart, which is a worse failure of the same kind.
• The benchmark comparison fails if DeepSeek publishes harness and effort settings for either its 10 September or its 13 August figures and they differ materially, or if it states which HLE subset the V4-Pro 42.7 was measured on and it is not the one this card assumes. This is the limb most likely to fail, and it is named because the HLE result is the only number on this card that cuts against the vendor's own framing and therefore the one a reader will lean on hardest.
• The generalisation — that a vendor deprecation date should be treated as a variable fails if no second frontier vendor announces and then withdraws a model deprecation within twelve months of this card's date. One reversal is one reversal; a case-insensitive grep of content/signals/*.md for deprecat, retire and phasing out, re-run after this card was finished, returns nine files and none of the eight besides this one records a second vendor withdrawing an announced model deprecation. The horizon is stated so the limb can actually close.

Provenance (what I actually read). Three documents, all fetched from the main thread on 15 September 2026: api-docs.deepseek.com/updates (the Change Log, whose og:description metadata reads "Date: 2026-09-10", matching its newest entry heading); api-docs.deepseek.com/quick_start/pricing (which carries no publication date of any kind on its face — the reason the date field above comes from the changelog); and api-docs.deepseek.com/news/news260910, the api-docs release announcement, whose page title is "DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient" and which carries no visible dateline of its own, though the site's own News index — read only as a listing, and a fourth page — lists it as "DeepSeek-V4.1-Flash Release 2026/09/10".

Every quotation attributed to a primary on this card comes from one of those three pages, with the single exception of the News-index listing string named above, enumerated: from the changelog — the "Date: 2026-09-10" heading, the reversal sentence, the V4-Flash retirement sentence, the V4.1-Flash benchmark list with its HLE footnote, and from its "Date: 2026-08-13" entry the V4-Pro benchmark figures; from the pricing page — footnote (2), footnote (1) on the legacy model names, the two model rows with their version strings and six quoted prices, the two concurrency limits, the off-peak sentence and the peak-hours definition; from the announcement page — its title, the "Tests by multiple parties" sentence, the "phasing out V4-Pro" sentence, the 04:00 UTC routing sentence, the "until V4.1-Pro launches" clause, and the "New pricing takes effect at 04:00 UTC on Sept 10, 2026" sentence. The changelog's 31 July and 21 August entries were read for whether they state an evaluation harness, and the fact that they do is asserted from that reading; no sentence of either is quoted.

Quotations from other cabinet cards, seven distinct quoted strings from four cards: three whole falsifier limbs from deepseek-v4-1-flash-routes-v4-pro-traffic (routing, cost, superiority); two from inference-cost-collapse (its Stanford HAI finding and, separately, its falsifier); one from perishable-frontier-model-shelf-life-weeks (its thesis with its self-label attached); and one from benchmark-number-one-means-one-board (its closing instruction, with its hype tier named). The earlier card's capsule fragment about the unread charts is quoted in that card's correction note, not here. Each is named inline at the point of use and comes from those files, not from any DeepSeek page.

Not opened and not sourced to: deepseek.com/en/news/deepseek-v4-1-flash/ (the page the earlier card cites, which is a separate URL from the api-docs announcement read here and was not re-opened, so this card makes no claim about whether that page has changed), the Hugging Face model card, the technical report PDF, the benchmark and price images on the announcement page, and the platform console. Two aggregator articles asserting the reversal surfaced this item and are recorded as leads only; neither contributed a fact, and the reversal is quoted here from DeepSeek's own pages.