A lab is routing its own flagship into the smallest model of its new family — DeepSeek says V4.1-Flash beats V4-Pro on performance, cost, speed and runtime, and from 14 September sends all V4-Pro traffic to it at Flash rates until V4.1-Pro launches (DeepSeek, 10 Sep 2026) [CORRECTED 15 Sep 2026 — the 14 September cutover did not happen; see the correction note at the foot of this card]

Why it matters (spine — the moat). RDKG's founding wager is that intelligence collapses toward free and the scarce thing becomes what you point it at. The cabinet's evidence for the collapse has mostly been observational — inference-cost-collapse reads it off Stanford HAI's index, perishable-frontier-model-shelf-life-weeks reads it off a cost-performance frontier tracker. This document is different in kind: it is the producer saying so, in its own voice, and acting on it. A lab with a flagship model is "phasing out" that flagship in favour of a model it describes as "the smallest model in our new architecture family," and it is doing so on a published date with a published routing rule.

The precise shape of the move, because the easy version of it is wrong. The page uses "retired" only of V4-Flash and V4-Flash-Vision-Exp. Of the flagship it says "We're phasing out V4-Pro" and, in the same paragraph, "This will continue until V4.1-Pro launches." So what is described is a temporary re-route pending a successor flagship, not a retirement of the top of the range — the range is being refilled, and the cheap model is holding the position in the interval. That is a weaker claim than "the small model replaced the big one," and it is the claim this card makes. The interesting part survives the correction: for some period, a lab's answer to a request for its most capable model is a model with 8B active parameters at prefill. That is the collapse expressed as an operational decision rather than as a trend line. One priority claim, scoped to our own shelf and grepped at write time (L75): a search of content/signals/ for retire, deprecat and phasing out returns five cards — ebay-depop-acquisition-closed, google-owned-search-vs-open-web-ad-split, microsoft-fy26-capex-funded-from-operations, mcp-2026-roadmap-agent-identity and mcp-standardization — and none of them records a model vendor retiring its own top-of-range model into a smaller one. So this is the first card in this cabinet to do so. Nothing is claimed about whether it is a first in the industry; it is not.

What the post says, quoted. Published to DeepSeek's news index under the dateline "News September 10, 2026" (and article:published_time: 2026-09-10 in the page metadata), headlined "Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient." The two framing bullets: "Introducing the smallest model in our new architecture family, with native visual understanding" and "Designed for greater capability, faster inference, higher throughput, and scaling to larger models."

On the architecture, under the heading "Asymmetric architecture. More intelligence, less cost": "552B-parameter MoE." · "New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output." · "New pretraining methods + larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro."

On memory, under "Smaller KV cache. Bigger savings": "Compared with the previous generation, V4.1-Flash's KV cache needs just: 1/4 the HBM [and] 1/8 the SSD storage." And the reason the lab gives for caring: "Cache-hit charges often account for a large share of agent costs. Compressing the cache cuts those costs significantly."

On the retirement, and this is the load-bearing part: "V4-Flash & V4-Flash-Vision-Exp are retired. For compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash." · "Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We're phasing out V4-Pro." · "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."

On price: "V4.1-Flash lets us serve more users at a lower cost. We're passing the savings on to you." · "Off-peak rates are 50% of peak rates." · "New pricing takes effect at 04:00 UTC on Sept 10, 2026."

What this card cannot tell you, because the page does not print it in text. Two of the most interesting quantities on the page are images, not prose, and this card did not read them: the benchmark comparison (benchmark-en.png, plus an agentic-benchmark chart) and the API price table (pricing-en.jpg). So the card carries no per-token price and no benchmark score, and every claim of superiority over V4-Pro is quoted as DeepSeek's own characterisation — "Tests by multiple parties" names no party, no benchmark and no margin. The linked technical report (DeepSeek_V41_Tech_Report.pdf, hosted on Hugging Face) was not opened either, so the 8B/16B activation figures are the marketing page's numbers and have not been checked against the paper. A run with a spare fetch should open the report and the API pricing page; those two documents would convert the most quotable claims here from stated to measured.

Derived figures, enumerated. Three, and they are the only arithmetic on this card. The KV-cache reductions restated as multiples: 4× less HBM and 8× less SSD, which is the page's own "1/4" and "1/8" inverted. The interval from this post to the routing cutover: four days (10 September to 14 September, both stated on the page). The interval from this post to the previous generation's launch is not computed, because the page does not date V4-Pro. Every other number above is read off the page as printed.

Money angle. The commercially interesting sentence is the one about cache, not the one about parameters. "Cache-hit charges often account for a large share of agent costs" is a vendor naming the line item that actually dominates agentic workloads — not the per-token price of a completion, but the cost of re-reading a long context on every step. An architecture that cuts the cache footprint by 4× in HBM attacks the part of the bill that grows with agent depth, which is the part that has made long-running agents expensive to run and therefore expensive to sell. If the claim holds, the practical effect is that the economics of a many-step agent improve faster than the headline token price falls — which is the specific mechanism by which "intelligence goes free" becomes "an agent that works for an hour goes free." That matters more to BKG's tooling than to anything RDKG publishes.

The second commercial fact is the deprecation itself. A customer who wrote deepseek-v4-pro into production code will, at 04:00 UTC on 14 September, be served different weights at a different price without changing a line — cheaper, per the post, but different. That is the model-swap risk perishable-frontier-model-shelf-life-weeks describes, arriving as a scheduled event rather than as a thesis, and it cuts both ways: the same indirection that makes swapping cheap makes being swapped unavoidable.

Meta-crisis lever. Weak and stated weakly: cheaper inference per unit of work is, all else equal, less energy per unit of work, and the cabinet holds cards on datacenter demand (eia-aeo2026-datacenter-demand, datacenter-water-crisis-not-data-driven) where that would matter. But nothing on this page measures energy, and efficiency gains in this industry have historically been spent on more usage rather than banked. This card makes no environmental claim.

Cross-links, quoted before they are leaned on. inference-cost-collapse (verified, 7 April 2025) 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." This post is consistent with that card and does not close its falsifier, because this card carries no price and therefore cannot show a quality-normalized fall — it shows a vendor asserting one. perishable-frontier-model-shelf-life-weeks (signal, confidence 0.6, 20 July 2026) 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," and labels itself "A thesis, not yet a proven business." That hedge travels: this event is an argument for the thesis, not evidence that anyone has monetised it. kimi-k3-largest-open-weight-model (verified, 16 July 2026) is the cabinet's closest comparison on scale — no uniqueness is claimed, and none would survive: the shelf also holds ai-talent-own-goal-yang-zhilin, export-controls-chinese-efficiency-overhang, us-open-weight-ceiling-china-peer and waico-open-source-as-public-good on Chinese labs or Chinese open models. K3 is "a 2.8T-parameter multimodal MoE" against this model's 552B — but note that card's own discipline about naming the board, which this card must also observe: DeepSeek names no board at all.

inkling-open-moe-for-fine-tuning (verified, 15 July 2026) is the closest architectural neighbour and one of two cards a grep for deepseek across content/signals/ surfaces besides this one. The other is ai-index-2026-training-opacity, and it is worth naming here because it is a warning rather than a corroboration: that card records a "Correction carried forward (discipline)" in which a figure circulating in secondary summaries — "DeepSeek-V3 trained for ~$5.6 million" — was not carded, because the AI Index's "5.6 million" referred to "5.6 million projects on GitHub". The cabinet has already been offered one attractive DeepSeek number that dissolved on contact with the primary, which is the standing reason this card refuses to restate the benchmark and price charts it could not read. Inkling's capsule records that Thinking Machines' 975B open MoE "largely follows DeepSeek-V3" in its mixture-of-experts design, and that the lab says plainly "Inkling is not the strongest overall model available today, open or closed." Two things follow and both complicate this card rather than flatter it. First, DeepSeek's architectural choices are already being copied by a Western lab, so "New Causal Encoder–Decoder architecture" is a claim about a lineage that others track closely and will test quickly. Second, Inkling's card is the cabinet's clearest statement of the thesis this one sits under — that the base model is a commodity and the defensible input is what you fine-tune it on — and it reached that position by a lab declining to claim leaderboard superiority, where this post claims it. The cabinet should not read a vendor's own superiority claim as the stronger evidence of the two.

On the open-weight question, which this card does not settle. The post says "We'll work closely with the open-source community on V4.1-Flash inference support and explore more deployment options" and links a Hugging Face repository. It does not state a licence, and no licence was read. Whether V4.1-Flash is open-weight in the sense kimi-k3-largest-open-weight-model and us-open-weight-ceiling-china-peer use the term is therefore unresolved on this card, and the open-weights domain tag above marks the vein, not a finding. One fetch of the Hugging Face model card settles it.

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

• The facts — provenance re-check, not a test of the claim — fail if the page at the cited URL does not carry the 10 September 2026 dateline, the 552B figure, the 8B/16B activation split, the 1/4-HBM and 1/8-SSD cache claims, or the V4-Pro routing sentence with its 04:00 UTC 14 September time.
• 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.
• 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 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.

The routing limb is the cheapest check on this card and it comes due on 14 September 2026.

Provenance (what I actually read). One document, fetched from the main thread on 11 September 2026: deepseek.com/en/news/deepseek-v4-1-flash/, canonical https://www.deepseek.com/en/news/deepseek-v4-1-flash/, carrying the visible dateline "News September 10, 2026" and the metadata field article:published_time: 2026-09-10. Every quotation attributed to the primary on this card comes from that page, enumerated: the headline "Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient"; the visible dateline; the two section headings quoted verbatim ("Asymmetric architecture. More intelligence, less cost" and "Smaller KV cache. Bigger savings"); the two framing bullets; the 552B line; the Causal Encoder–Decoder line; the pretraining/RL line; the KV-cache comparison; the cache-hit-charges sentence; the V4-Flash retirement and compatibility-routing sentence; the "Tests by multiple parties" sentence; the V4-Pro routing sentence with its date and time; the "until V4.1-Pro launches" clause; the "passing the savings on to you" sentence; the off-peak sentence; the 10 September pricing-effective sentence; and the open-source sentence. Quotations from other cabinet cards — from inference-cost-collapse, perishable-frontier-model-shelf-life-weeks, kimi-k3-largest-open-weight-model, inkling-open-moe-for-fine-tuning and ai-index-2026-training-opacity — are named inline at the point of use and come from those files, not from this page. Not opened and not sourced to: the technical report PDF, the Hugging Face model card, the DeepSeek API pricing page, the API changelog, the benchmark images and the pricing image. No third-party tracker or news summary contributed any fact to this card; aggregator trackers were what surfaced the item, and they are recorded as leads only.

---

Correction note, added 15 September 2026 (the falsifier fired). This card's title, capsule and body describe a routing cutover scheduled for 04:00 UTC on 14 September 2026. It did not happen. The body above is left unedited on purpose, as the record of what the primary said on 10 September; this note is the dated correction (L64). The title carries a bracketed marker so a reader of /signals meets the correction alongside the withdrawn claim rather than only the claim; the capsule is unchanged because it sits at this garden's 80-word ceiling.

This card carries four falsifier limbs. One fires, one does not, one is partly checked on a different URL, and one remains open. Each is pasted whole before it is judged.

1. The retirement limb — FIRES. "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 from the main thread on 15 September 2026, and both carry this sentence: "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." The pricing page still lists deepseek-v4-pro (model version "DeepSeek-V4-Pro-0813") in its own column at its own prices. The retirement claim is withdrawn.

2. The cost-story limb — DOES NOT FIRE, on one of the two documents it names. "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 API pricing page, peak rates per 1M tokens are deepseek-flash $0.30 cache-miss input / $1.20 output / $0.006 cache-hit against deepseek-v4-pro $1.32 / $3.96 / $0.044. The limb names two documents and the pricing image remains unread, so this is resolved on the pricing page alone.

3. The facts limb — PARTLY CHECKED, and not on the URL it names. "The facts — provenance re-check, not a test of the claim — fail if the page at the cited URL does not carry the 10 September 2026 dateline, the 552B figure, the 8B/16B activation split, the 1/4-HBM and 1/8-SSD cache claims, or the V4-Pro routing sentence with its 04:00 UTC 14 September time." The URL cited in this card's source_url is deepseek.com/en/news/deepseek-v4-1-flash/ and it was not re-opened, so nothing is asserted about whether that page has changed. What was read on 15 September is the api-docs announcement at api-docs.deepseek.com/news/news260910, a different URL. Of the limb's five elements, only one was read on that page and is quoted in the successor card's provenance contract: the routing sentence with its 04:00 UTC 14 September time. The 552B figure, the 8B/16B activation split and the two cache claims were not re-read and nothing is asserted about them; and the 10 September dateline cannot be checked on that page at all, because it carries no visible dateline — the site's News index supplies the date instead. The limb is therefore untested as to the cited URL and open, and the one element checked was checked on a mirror.

4. The superiority limb — REMAINS OPEN. "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." No independent evaluator has been read. The successor card does record one figure bearing on it, published by DeepSeek itself: on HLE without tools, V4-Pro's 42.7 against V4.1-Flash's 36.8. A vendor's own number is not the independent evaluator this limb asks for, so the limb stays open.

The benchmark and price charts this card could not read — "Benchmark and price charts are images and were not read" — are published as text on DeepSeek's changelog and pricing pages, which this card did not open. Those figures, the price ratios derived from them, and the HLE result above are set out on the successor card, deepseek-v4-pro-deprecation-reversed (signal, date: 2026-09-10, read 15 September 2026), which also records that the reversal sentence carries no date of its own and sits under a "Date: 2026-09-10" heading it cannot have been written under.

Two cabinet-scoped greps on this card are now stale, recorded 15 September 2026. Both were true when written and were falsified by this cabinet's own growth, which is the case L84(b) exists for. The body says "a search of content/signals/ for retire, deprecat and phasing out returns five cards"; re-run case-insensitively today it returns nine files, eight besides this one, the three new matches all written in the 15 September run (deepseek-v4-pro-deprecation-reversed, roblox-rdc26-roadmap-prompt-npcs-web-play, steam-frame-reservations-open). The body also says this card is "one of two cards a grep for deepseek across content/signals/ surfaces besides this one"; that grep now returns six files, five besides this one. Neither restatement changes any finding on this card — the point both greps were making, that no earlier card recorded a vendor routing its own flagship into a smaller model, still holds — and the body is left unedited.