DeepSeek is retiring its flagship by routing it away — and on 14 September anyone still calling V4-Pro gets a reasoning downgrade with the price cut
TL;DR: DeepSeek shipped V4.1-Flash on 10 September 2026 — a 552B MoE activating only 8B per input token / 16B per output token, MIT-licensed weights on Hugging Face, KV cache down to 890 bytes/token. Off-peak pricing of $0.15 in / $0.60 out per million (peak exactly 2x) is the cheapest credible frontier-adjacent tier available. The item that needs a calendar entry is four days out: from noon Beijing / 04:00 UTC on 14 September, requests to deepseek-v4-pro route to Flash and bill at Flash rates, until a V4.1 Pro with no announced date. That is a 70–86% price cut — and a capability trade. Flash beats Pro on agents (DeepSWE 74.2 vs 62.7, Terminal-Bench 90.6 vs 87.9) and loses on reasoning (GPQA Diamond 90.9 vs 92.4, HLE 39.1 vs 42.7). Nothing in your code breaks; the model behind your endpoint just changes. Separately, the peak windows (01:00–04:00 and 06:00–10:00 UTC, Mon–Fri) put European mornings at double rates while the entire US working day sits off-peak. For buyers: pin explicit model IDs before Saturday, and if you are in Europe, move batch work past midday local.
Update (12 September 2026): Anthropic’s threat report of 10 September alleges that DeepSeek silently relayed customer exchanges to Claude — more than 12.1 million of them over a two-week window in July 2026 — without notifying those customers. The allegation is vendor-authored and DeepSeek has not responded publicly. It is worth reading alongside the routing change below for the same reason: in both cases the model actually answering your request is decided by the vendor, and your evaluation harness cannot see it.
The launch is not the news
The model itself is impressive and roughly what the market expected. DeepSeek has spent 2026 pushing the floor down — we covered it when it bolted vision onto V4-Flash and charged nothing extra, and the pattern has held through GLM-5.3 Flash and Qwen-Flash-Next and Gemini 3.8 Flash.
Here is what launched:
| Flash off-peak | Flash peak | Pro off-peak | Pro peak | |
|---|---|---|---|---|
| Cached input | $0.003 | $0.006 | $0.022 | $0.044 |
| Uncached input | $0.15 | $0.30 | $0.66 | $1.32 |
| Output | $0.60 | $1.20 | $1.98 | $3.96 |
Per million tokens. Peak is exactly 2x off-peak on every line.
Context is 1M tokens with a maximum output allowance of 384K, and Flash carries a concurrency limit of 2,500. Architecturally it is a 552B MoE backbone with about 196B of conditional Engram memory, activating ~8B parameters per input token and ~16B per output token through an asymmetric causal encoder-decoder — a real departure from decoder-only convention — with compressed sparse attention, n-gram lookups, a vision encoder and speculative decoding.
The number that funds the price list is the cache: 890 bytes of global KV cache per token, against 3,514 for V4 Flash — about a 3.9x reduction, which DeepSeek reports as a quarter of the HBM and an eighth of the SSD footprint of the previous generation. Cheap cache is why cache-hit input can be priced at a third of a cent per million.
The part that needs a calendar entry
From noon Beijing time on 14 September 2026 — 04:00 UTC — requests to
deepseek-v4-proare routed to V4.1-Flash and billed at Flash prices, until V4.1 Pro ships. No date has been given for V4.1 Pro.
DeepSeek is not deprecating its flagship with a sunset notice and a migration window. It is routing the flagship away and leaving the identifier live. Your calls keep succeeding. Your latency probably improves. Your bill falls by roughly three quarters. Nothing fails, which is exactly why this is easy to miss.
The price delta on the routing change:
| Line | V4-Pro peak | V4.1-Flash peak | Cut |
|---|---|---|---|
| Output | $3.96 | $1.20 | −70% |
| Uncached input | $1.32 | $0.30 | −77% |
| Cached input | $0.044 | $0.006 | −86% |
A 70-86% cut arriving unrequested should prompt one question: what is being traded for it?
It is a capability trade, and it cuts both ways
The benchmarks answer that, and the answer is not uniform.
| Benchmark | V4.1-Flash | V4-Pro | Winner |
|---|---|---|---|
| DeepSWE | 74.2 | 62.7 | Flash +11.5 |
| Terminal-Bench 2.1 | 90.6 | 87.9 | Flash +2.7 |
| GPQA Diamond | 90.9 | 92.4 | Pro +1.5 |
| HLE (text-only) | 39.1 | 42.7 | Pro +3.6 |
Read that against the architecture and it stops being surprising. A model activating 8B parameters on input and 16B on output, tuned around agent workload shapes, should do well at long-horizon tool use and less well at dense single-shot reasoning. That is what the table shows: Flash is a better agent and a worse thinker.
So the 14 September routing lands very differently depending on why you chose Pro:
- Agent and coding pipelines get an 11.5-point DeepSWE improvement and a ~70% price cut. Unambiguously good. Do nothing except confirm it.
- Analysis, research, and hard reasoning workloads lose 1.5 points of GPQA Diamond and 3.6 of HLE, on the precise axis Pro was selected for — and there is no paid tier to escape to, because Pro is gone until V4.1 Pro appears.
That second group is the one that needs to act, and the discount is what will stop them noticing. A bill that falls does not generate a ticket.
The peak-hour windows are geography, not just timing
The peak/off-peak split is exactly 2x on every line, which makes it one of the largest single levers on this price sheet. Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — seven hours a weekday, 35 of 168, so about 79% of the week is off-peak.
Now place those windows on a map. In September, with CEST at UTC+2 and US Eastern at UTC−4:
| Location | 06:00–10:00 UTC block lands at | Effect |
|---|---|---|
| Frankfurt / Paris (CEST) | 08:00–12:00 | Entire working morning at 2x |
| London (BST) | 07:00–11:00 | Most of the morning at 2x |
| New York (EDT) | 02:00–06:00 | Overnight — off-peak all workday |
| San Francisco (PDT) | 23:00–03:00 | Overnight — off-peak all workday |
| Bengaluru (IST) | 11:30–15:30 | Straight through midday |
| Beijing (CST) | 14:00–18:00 | Afternoon at 2x |
The windows track Chinese, Indian and European demand — which is rational capacity pricing for DeepSeek’s actual load curve. The consequence for buyers is that an American team pays off-peak rates for its entire working day, and a European team pays double for its entire morning, for identical work on an identical model.
That is a real and rarely-stated cost asymmetry, and it is the mirror image of the problem we described when Mistral began selling the only priced in-region priority tier for EU buyers. European teams pay a premium for in-region guarantees from EU vendors and a premium for working hours from Chinese ones.
The mitigation is cheap: for anything batch — evaluations, backfills, nightly agent runs, document processing — schedule past 12:00 CEST and the line halves. Interactive traffic you cannot move, but interactive traffic is rarely the volume.
What this says about the “flagship” tier generally
Three vendors have now demonstrated that the flagship is the least stable thing on a price sheet. OpenAI’s Astra re-bills an entire request at premium rates past a 272K input-token cliff. Qwen3.8-Max-0902 shipped better coding performance at an unchanged price. And DeepSeek has now removed its flagship from service by routing rather than retirement.
The through-line is that the name you call is not a stable description of what you get. Model identifiers are increasingly aliases pointing at whatever the vendor currently wants to serve, and the cost of that indirection is that capability can move underneath a workload without any error, any email, or any line in a changelog. It is a milder, commercial cousin of the concern raised by joint advisory AA26-251A’s recommendation that providers serve varied degraded responses to suspected distillers — an advisory that, two days before this launch, named DeepSeek specifically.
The defence is the same in both cases and it is not exotic: a small private canary set, run on a schedule against every endpoint you depend on, scored automatically, tracked over time. Thirty to fifty fixed prompts with known-good outputs would show a GPQA-shaped regression on 14 September within a day. Without one, you will find out when a customer does.
It is also worth noting what these prices do to the arithmetic in our flagship piece on OpenAI’s 10,000-agent swarm. That run implied roughly $11 per agent-hour at Astra rates. The same sustained throughput on V4.1-Flash off-peak would run about $0.09 per agent-hour on output — over two orders of magnitude less. Nobody is claiming Flash could have done that work. But it does mean the cost of trying an agent architecture and the cost of running the frontier are now separated by roughly 100x, and most teams should be prototyping on the cheap end of that gap. The full per-agent-hour table across eight models is in best AI agents.
What to do before Saturday
- Inventory every call to
deepseek-v4-proand pin an explicit model identifier instead of the alias. - Sort those workloads into agent-shaped and reasoning-shaped. The first group benefits; the second regresses.
- For reasoning-shaped work, choose a destination anyway — Pro survived the 14th (see the update at the foot of this piece), but on a reprieve DeepSeek can withdraw with a changelog line. Our DeepSeek V4-Pro versus Claude Opus 4.8 comparison covers the obvious alternative.
- If you are in Europe, move batch work past 12:00 local. It halves that line, immediately, for no engineering effort.
- Stand up a canary set before the routing change, so you have a before-and-after rather than a hunch.
The DeepSeek review tracks the model line and its pricing history, and our best AI coding tools rundown covers where a 74.2 DeepSWE score at $0.60 per million output tokens actually fits.
Update, 11 September 2026 — the same unobservable surface, sold as the product. The argument above is that a routing decision you cannot see is a capability change you cannot evaluate, and that a canary set is the only defence. Sakana AI shipped the maximal version of that trade on 11 September: Fugu Max and Fugu Ultra v2 are orchestrators whose entire value is choosing which model answers, and whose documentation states that the models selected and how they are coordinated are proprietary and not exposed by design. Unlike DeepSeek’s change, nothing here is a downgrade and no capability is being withdrawn — but the buyer’s position is identical. You get token totals without attribution, a veto over specific providers in console settings, and no way to tell a cheaper worker from a good day. The canary-set advice in point 5 above applies unchanged, and with more force: run a fixed prompt set on a schedule, because an undisclosed routing policy changes without a changelog.
Update, 14 September 2026 — the retirement was reversed, which is its own finding. The deadline in this article passed without the outcome it describes. DeepSeek’s API changelog now states that “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,” and that it “will provide further notice should there be any changes.” The V4.1-Flash release notes have not been amended — they still say all deepseek-v4-pro requests route to Flash from 04:00 UTC on 14 September — so a buyer reading DeepSeek’s own documentation today finds two statements that contradict each other, with no stated precedence and no version history. The prose above has been corrected where it asserted Pro would be unavailable at any price; that is no longer true. Everything else in the piece stands, and the central argument is strengthened rather than weakened. A model line whose availability can be withdrawn on four days’ notice and then restored on the day, with the superseded notice left in place, is not a planning surface. The reprieve has no end date attached and was granted at the vendor’s discretion, so the action items are unchanged: pin explicit model identifiers rather than aliases, keep the canary set running so you can tell which model answered regardless of what the docs say, and keep a tested destination for reasoning-shaped work. The same week produced a mirror-image case at the other end of the market: Anthropic’s subscription limits moved in the opposite direction on the same date, for reasons that also had nothing to do with capacity — see the 80% inference margin and the Claude Code limit cut.
Update, 26 September 2026 — the same failure mode, in the opposite direction. This article’s argument was that capability can move under a stable model name without any rate-card or version-string change. September supplied the mirror image. On 25 September OpenAI fixed “a bug in image encoding that degraded image understanding” in GPT-6 Sol and GPT-6 Luna, three days after launch — a change that moved capability up behind an unchanged identifier, with no dated snapshot available to pin either build. And on 23 September Fireworks released Ember-1 at exactly base Kimi K3’s rate card, where the entire claimed saving is fewer tokens consumed rather than a lower price per token. Three cases, one conclusion: the two variables buyers actually monitor — unit price and version string — are the two that stayed still. The metric that would have caught all three is tokens per completed task, logged per model, per day. How the unversioned-model problem plays out at OpenAI.
Frequently asked questions
What exactly changes on 14 September, and what do I have to do?
From noon Beijing time on 14 September 2026 — 04:00 UTC — requests sent to the deepseek-v4-pro model identifier are routed to V4.1-Flash and billed at Flash prices. This continues until a V4.1 Pro ships, and no date has been given for that. You do not have to change any code for your requests to keep succeeding, which is precisely the risk: nothing breaks, your bill falls sharply, and the model behind your endpoint quietly changes. The action item is to decide before the 14th whether Flash is acceptable for each workload currently pointed at Pro, and to pin an explicit model identifier rather than relying on the alias. Update, 14 September 2026: DeepSeek reversed this. Its API changelog now states that '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,' adding that it 'will provide further notice should there be any changes.' The V4.1-Flash release notes still carry the original routing statement, so the two notices sit side by side in DeepSeek's own documentation. Treat Pro as available and unretired, verify which model actually answers before relying on it, and keep the explicit model identifier pinned — the reprieve is discretionary and carries no end date.
Is the routing change a price cut or a downgrade?
Both, and which one dominates depends entirely on your workload. On price it is unambiguous: Flash undercuts Pro by about 70% on output ($1.20 versus $3.96 peak per million), 77% on uncached input ($0.30 versus $1.32), and 86% on cached input ($0.006 versus $0.044). On capability it splits. Flash is materially better on agent and coding work — DeepSWE 74.2 against Pro's 62.7, a gap of 11.5 points, and Terminal-Bench 2.1 at 90.6 against 87.9. But Pro remains ahead on pure reasoning: GPQA Diamond 92.4 versus 90.9, and HLE text-only 42.7 versus 39.1. So for agent pipelines this is a free upgrade that also cuts the bill by roughly three quarters. For analytical or research workloads chosen specifically for Pro's reasoning, it is a quiet regression on the exact axis you were paying for.
How do the peak and off-peak windows actually work, and do they favour some buyers?
Peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — seven hours a weekday, 35 hours out of 168, so about 79% of the week is off-peak. Off-peak is exactly half of peak on every single line item, with no exceptions. The geography is where it gets interesting. In September, with CEST at UTC+2, the 06:00–10:00 UTC block lands at 08:00–12:00 in Frankfurt or Paris — the European working morning, charged at double. The same block is 02:00–06:00 US Eastern and 23:00–03:00 US Pacific, meaning an American team's entire working day sits in off-peak. Indian teams see 11:30–15:30 IST, straight through midday. The windows track Chinese, Indian and European load, and US buyers get a structural 50% discount by accident of longitude. If you are in Europe and running batch work, shifting it past 12:00 local halves that line of your bill.
What is different about the architecture, and does it matter commercially?
V4.1-Flash is a 552B mixture-of-experts backbone plus roughly 196B of conditional Engram memory, but it activates only about 8B parameters per input token and 16B per output token — an asymmetric causal encoder-decoder rather than the standard decoder-only design, with compressed sparse attention, n-gram lookups, a vision encoder and speculative decoding. The commercially relevant consequence is memory: global KV cache falls to 890 bytes per token against 3,514 for the previous V4 Flash, roughly a 3.9x reduction, which DeepSeek reports as a quarter of the HBM and an eighth of the SSD footprint of the prior generation. That is what funds the cache-hit price of $0.003 per million off-peak, and it is why the asymmetry — 8B active on input, 16B on output — is described as matching agent workload shapes, where enormous contexts are read and comparatively little is written.
The weights are MIT-licensed on Hugging Face. Should I self-host instead?
Probably not yet, and DeepSeek effectively says so. The published weights — around 485B parameters in a mixed BF16/FP8/INT8 format — are labelled preliminary and intended for community inference rather than production deployment. Beyond that caveat, the economics are brutal at these prices: at $0.60 per million output tokens off-peak, you would need enormous sustained volume before self-hosted GPUs beat the API on total cost, and you would forgo the speculative decoding and cache infrastructure that make the hosted numbers achievable. The genuine case for taking the weights is not cost but continuity and jurisdiction — having a tested fallback you control if the hosted endpoint becomes unavailable, gets repriced, or becomes procurement-problematic. Given that DeepSeek was named in joint advisory AA26-251A two days before this launch, that second consideration is not hypothetical for every buyer.
Does the CISA advisory naming DeepSeek change whether I should use this model?
It changes the procurement question more than the technical one. Advisory AA26-251A, issued 8 September by the NSA, CISA and FBI, names DeepSeek among six Chinese firms alleged to have distilled US frontier models at industrial scale. Nothing in it prohibits using DeepSeek's models or downloading its weights, and running open weights you obtained legally is not what the advisory targets. But there is now a documented federal position naming this specific vendor, which raises the odds of future export or Entity List action touching the hosted API, and which matters if your customer contracts, public-sector exposure or supply-chain attestations are sensitive to provenance. The proportionate response is the same as before: keep a routing abstraction, keep a tested alternative, and do not put anything load-bearing behind a single named vendor's hosted endpoint — advice that the 14 September routing change independently validates.
Sources
- DigitalApplied — DeepSeek V4.1 Flash: Pro Routing, Prices and Early Tests
- DeepSeek API Docs — Change Log (V4 Pro continuity notice, September 2026)
- DeepSeek API Docs — DeepSeek-V4.1-Flash release notes (10 September 2026)
- PANews — DeepSeek V4 Pro API Call Service Will Not Be Discontinued (September 2026)
- AlphaSignal — DeepSeek's V4.1-Flash Beats Its Own Flagship at a Quarter of the Memory Cost
- CellCog — DeepSeek V4.1 Flash Release Date, With New Flash Pricing From the Same Day
- ccleaks — DeepSeek V4.1 Flash open weights land on Hugging Face
- AIToolsReview — DeepSeek V4.1 Flash: Specs, Pricing & Review (September 2026)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.