DeepSeek V4.1 Pro Is Still Unreleased: A Cancelled Reroute, a Shipped Flash, and a Tracker's Diagnosis

A week ago, DeepSeek V4.1 Pro was the name at the end of a routing notice. Today it is the name at the end of a routing notice the company cancelled. As of September 16, 2026, DeepSeek V4.1 Pro is still unreleased — no model card, no endpoint, no price, no date — while DeepSeek V4.1 Flash, the model meant to cover the gap in the meantime, shipped on September 10 as an open-weights release, and the plan to retire DeepSeek V4 Pro to make room for the successor was called off before its September 14 deadline. The family has moved on one axis and stalled on another, and the explanation for the stall that has travelled furthest is an unfashionable one: swarm coordination.

The September 9 notice, for the record, said this. Tianyi Cui, a member of DeepSeek's technical team posting as @tianyi, wrote that once DeepSeek V4.1 Flash went official, all API requests aimed at DeepSeek V4 Pro would be rerouted to it and billed at Flash rates, and that the arrangement would hold until DeepSeek V4.1 Pro itself launched. The first half of that plan happened. The second half did not: after users objected to having a production model swapped underneath them without a migration window, DeepSeek reversed itself and said DeepSeek V4 Pro would keep being served, billing unchanged. The clause that survived the week was the one nobody was meant to notice — that DeepSeek V4.1 Pro exists.

One sourcing note before the details, because the evidence here is deliberately uneven. Everything about DeepSeek V4.1 Pro still rests on inference: a company statement that names it, and one tracker's reading of what is holding it up. Everything about DeepSeek V4.1 Flash is now vendor documentation — a changelog entry, a model card, published weights, a price sheet. The reversal on DeepSeek V4 Pro is on the record in DeepSeek's own API docs. And the explanation for the delay has one source and no corroboration at all. This piece keeps those tiers apart, because an announced plan, a withdrawn plan, a shipped model and a commentator's diagnosis are four different kinds of evidence.

What happened in the week after the notice

The September 9 notice described an intention. Within 48 hours it had been overtaken twice.

On September 10, DeepSeek shipped DeepSeek V4.1 Flash — the model that had been running as a two-day beta under the throwaway ID deepseek-v4.1-flash-expires-on-0910. It is a generational change rather than a point release: a 552-billion-parameter mixture-of-experts model on a new Causal Encoder-Decoder architecture, with 8 billion parameters active during prefill and 16 billion during decode, a one-million-token context and up to 384K tokens of output. Vision is built into pre-training rather than bolted on as a separate encoder, which is what the "native multimodal" label is doing. DeepSeek's changelog frames it as "the smallest model in our new architecture family, with native multimodal visual understanding."

It also shipped with open weights, which closes a question the September 9 version of this story left open. The repository deepseek-ai/DeepSeek-V4.1-Flash went up on Hugging Face the same day under the MIT licence and has since passed 325,000 downloads. A DeepSeek staff member, Zizheng Pan, described the release as "our first flagship model with native multimodal support" — worth quoting as staff framing, though DeepSeek's own release note positions the model as the smallest member of the new family rather than its flagship. Both readings come from the company; the weights are the part anyone can check.

The pricing is DeepSeek's own and it is steep. For the model now callable as deepseek-flash, input runs $0.30 per million tokens at peak and $0.15 off-peak, and output $1.20 at peak and $0.60 off-peak. On the same sheet DeepSeek V4 Pro is $1.32 input and $3.96 output at peak — output on the new model is roughly 70% cheaper than on the tier it was briefly meant to replace, which is the number that made the reroute plan look reasonable in the first place. Independent testing has since added one caveat: Artificial Analysis notes that DeepSeek V4.1 Flash consumes a relatively high number of output tokens per task, which eats into the headline saving once you are billed for the tokens actually generated rather than the ones advertised.

The second thing that happened was the reversal. DeepSeek had set the DeepSeek V4 Pro cutover for September 14 at noon Beijing time. Developers pushed back — the objection being that redirecting a production model to different weights breaks prompt-tuned pipelines and agentic workflows built against the old ones — and the company changed its mind. DeepSeek's API documentation now says it will "continue providing API services for DeepSeek V4 Pro after September 14, 2026," with "the billing method remaining unchanged," and will give notice of any change. It is worth noting what the retirement could and could not have done: DeepSeek V4 Pro's weights are published under MIT, so no API change could have made the model unavailable — only unavailable from DeepSeek. Several Chinese business outlets reported the reversal on September 11, ahead of the deadline it was meant to beat.

What it means if you call DeepSeek V4 Pro today

The practical answer as of September 16 is that nothing changes. A request naming DeepSeek V4 Pro is still answered by DeepSeek V4 Pro, still billed at its existing rates, with no migration to perform and no code to change. That is the opposite of what this page said a week ago, and it is worth being concrete about why: the redirect was announced, scheduled, objected to and withdrawn inside five days.

The retirement plan was not unreasonable on its own terms. DeepSeek's stated reasoning was that continuing to charge premium rates for DeepSeek V4 Pro while DeepSeek V4.1 Flash was faster, cheaper and stronger would be hard to defend — and the price gap backs that up. What the plan underestimated was that a server-side redirect is not a price cut, it is a model swap, and teams that had spent months tuning prompts and agent loops against DeepSeek V4 Pro were being asked to absorb it with roughly a day of notice.

The capture above is DeepSeek's Models & Pricing page from September 9, kept as a dated snapshot: at that point the lineup was the plain V4 generation — DeepSeek V4 Flash, DeepSeek V4 Pro and DeepSeek-V4-Flash-Vision-Exp — with no V4.1 entry anywhere in the table. The page has since been rewritten. It now lists two models, deepseek-flash and deepseek-v4-pro, and carries a footnote saying DeepSeek V4 Pro will keep being served after September 14 with billing unchanged. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted for compatibility, but the models behind them have been retired, so those requests are served by DeepSeek V4.1 Flash and billed at Flash rates.

Why the V4 paper is still the key to reading the leak

teortaxesTex's original claim was that DeepSeek V4.1 Pro being in the works confirms DeepSeek has solved architectural issues of the V4 family "as they committed to do in the V4 paper." That reference is checkable, and the week since has made it more interesting rather than less. DeepSeek's V4 technical report — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence — is unusually candid about what its design did not yet do well:

• The report calls its own architecture "relatively complex" and commits that "in future iterations, we will carry out more comprehensive and principled investigations to distill the architecture down to its most essential designs."

• Its two training-stabilization mechanisms — Anticipatory Routing and SwiGLU Clamping — worked in practice, the report says, but their underlying principles remained "insufficiently understood."

• Long-context retrieval still decays past 128K tokens. On an 8-needle retrieval test cited in the report, V4-Pro-Max scores roughly 0.92 at 128K and falls to about 0.59 at 1M — the gap between being able to load a million tokens and being able to use them.

• And the V4 line shipped text-only; vision arrived later as a bolted-on experimental encoder, DeepSeek-V4-Flash-Vision-Exp, rather than as part of the base model.

Read against that list, DeepSeek V4.1 Flash is the first shipped evidence that the V4 paper's promises were being acted on rather than merely repeated. The release describes a genuinely new architecture, "designed for a higher capability ceiling, faster inference, higher throughput" and intended to scale to larger models; images are taken in natively from pre-training instead of through a separate encoder; and DeepSeek puts the persistent KV-cache footprint at roughly an eighth of DeepSeek V4 Flash's and the per-token global cache at about a quarter — the long-context cost problem the V4 report had identified. None of that is a DeepSeek V4.1 Pro specification, and it does not settle the retrieval decay past 128K tokens, for which no V4.1 figure has been published. It is a strong hint about direction, not a guarantee about destination.

It is also why the inference should stay labelled as one. DeepSeek has now published a V4.1 base — Flash — but no V4.1 technical report, and nothing at all about DeepSeek V4.1 Pro's architecture. "In the works" remains supported only by a team member's reference to its eventual launch.

The tracker's diagnosis: swarm coordination, not scale

On September 10, teortaxesTex posted what amounts to a theory of the delay: "DeepSeek needs to complete V4.1-Pro, tune it better on swarm coordination, and solve some open problem. Maybe not a fancy one in research math (OpenAI will sweep it), but something applied that's been overlooked. I believe it can be done."

Be precise about what that is. It is not reporting, and it is not sourced to anyone inside DeepSeek: it is one observer's read, published on the day DeepSeek V4.1 Flash shipped, and no DeepSeek document names swarm coordination as a blocker. Nothing in the company's changelog, model card or price sheet mentions it. Treat it as a hypothesis about the delay offered by someone who has tracked this family closely — not as a fact about the roadmap.

It is, though, a coherent hypothesis, because it points at the workload DeepSeek V4.1 Flash is explicitly built for. The release pitches the model at input-heavy, long-context work — coding and research tasks where an agent reads far more than it writes — and DeepSeek's KV-cache numbers are a statement about the cost of holding many large contexts at once. "Swarm coordination" has no official definition here, but read plainly it describes the problem of keeping many agents working in parallel consistent with one another and stopping them from duplicating each other's work. That is the natural next problem once a single agent's context is cheap, and it is an applied engineering problem rather than a research-mathematics one — which is exactly the shape teortaxesTex describes.

If that reading is right, the delay has a shape worth naming: DeepSeek appears to have solved the part of the stack a benchmark can measure before solving the part that only shows up in production. The V4 report's own list of unfinished work is about architecture complexity, training stability, retrieval and multimodality; coordination between parallel agents is not on it. On this reading the bottleneck moved somewhere the company has never published a plan for — which is a reason to treat the diagnosis as interesting rather than established.

The evidence ledger, updated

Here is the ledger as it stood on September 9, and how each line reads now.

• Confirmed by vendor documentation — DeepSeek V4.1 Flash shipped on September 10, 2026, callable as deepseek-flash, with native image understanding, a 552B-parameter MoE backbone on a Causal Encoder-Decoder architecture, 8B and 16B active parameters at prefill and decode, a 1M-token context, and MIT-licensed weights on Hugging Face. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and rerouted to it.

• Confirmed by vendor documentation, and a reversal — the plan to retire DeepSeek V4 Pro on September 14 was announced and then withdrawn, and DeepSeek V4 Pro continues at unchanged billing. This is the line in the September 9 ledger that simply stopped being true within a week, in the reader's favour.

• Still open — DeepSeek V4.1 Pro's specs, parameter count, context window, pricing and launch date. Neither the changelog nor the pricing page names it, and it has no model card. Whether it ships open weights is genuinely unknown, but the default has shifted: the rest of the V4.1 line is now MIT-licensed, which makes an API-only Pro the less likely outcome rather than the expected one.

• One observer's hypothesis, uncorroborated — that the delay is about completing DeepSeek V4.1 Pro, tuning swarm coordination and solving an applied open problem. Single source, no vendor confirmation, no second reporter. Plausible, unverified, and the reason this piece keeps saying so.

• Interpretation, unchanged — that the V4.1 line proves the V4 paper's architectural commitments were kept. The base model that shipped is strong evidence of direction; it is not proof about a model nobody has seen.

What to watch

• A DeepSeek V4.1 Pro model card, or any DeepSeek document that names it. The changelog's most recent entry is September 10 and it is the V4.1 Flash release. A second entry is the event that would turn this page from a watch into a review.

• Whether the cancelled retirement stays cancelled. DeepSeek's note says it will give further notice of any change, so a second attempt at retiring DeepSeek V4 Pro is the thing to watch — and the next version of it will probably arrive with more notice than the last.

• First independent benchmarks of DeepSeek V4.1 Flash beyond the token-economics caveat — in particular whether the new base actually fixes the V4 generation's retrieval decay past 128K tokens, the gap DeepSeek's own report measured.

• Whether swarm coordination turns out to name a real blocker or was one observer's framing. If the next DeepSeek release note leads with agent coordination rather than benchmark scores, the diagnosis will have aged well.

What to do meanwhile

For anyone who spent the week bracing for a forced migration, the outcome is good news with a caveat. The good news: DeepSeek V4 Pro is still served at unchanged billing, so a workflow tuned against it does not have to move today and the cost of waiting is zero. The caveat: the successor to that tier is now unambiguously unfinished, and the explanation for the delay that has travelled furthest is a problem with no published plan behind it. The reasonable posture is the one this page recommended a week ago and which the week has only reinforced — keep the production path on the model you have already tuned, and treat the V4.1 line as something to test against rather than depend on.

That is exactly the posture a routing layer is built for. On OrcaRouter, DeepSeek V4.1 Flash and DeepSeek V4 Pro are both reachable through one API at DeepSeek's list price passed through with zero markup — the Flash and Pro rates on our model pages are DeepSeek's own numbers, not ours marked up — so a vendor-side price change or reroute lands on our side the same day it ships rather than whenever a rate sheet gets updated. Send a slice of traffic at DeepSeek V4.1 Flash to find out what it does to your token bill, keep DeepSeek V4 Pro as the stable path with automatic failover behind it, and let the routing DSL decide which calls are worth spending on an unproven model and which should stay on the workhorse. None of that needs a second contract, a second key, or a code change when the next V4.1 model eventually appears.

The bottom line on a stalled roadmap

The cleanest way to read the last week is that DeepSeek shipped the half of the V4.1 story that was ready and cancelled the part that would have forced anyone's hand. DeepSeek V4.1 Flash is real, open-weights and cheap, and it settles much of what the V4 paper promised about efficiency. DeepSeek V4 Pro survives a retirement that lasted five days, which is a reminder that redirecting a production model is a product decision rather than a billing one — and that a model with published MIT weights can be retired from an API but not from the world. And DeepSeek V4.1 Pro is still a name, now with a plausible account of what is holding it up, offered by one observer, corroborated by nobody and contradicted by nothing. Until DeepSeek publishes a card, a spec or a date, the rational position is unchanged: treat V4.1 Pro as real, treat every detail about it as unconfirmed, and treat the Flash release as the only part of this story you can actually download.

While DeepSeek V4.1 Pro has nothing to call yet, the current flagship you can reach through one API today is DeepSeek V4 Pro on OrcaRouter — at DeepSeek's list price, passed through with 0% markup.

And the model DeepSeek actually shipped, if you want to run your own token-economics test against it: DeepSeek V4.1 Flash on OrcaRouter.