DeepSeek V4.1 Pro vs DeepSeek V4 Pro: The Flagship You Can Call and the One You Cannot

The search that brings most readers here is a fair one: which DeepSeek flagship should I use. The honest answer has two halves, and they are not symmetrical. DeepSeek V4 Pro is a shipped model with a model card, a price sheet and an endpoint — 1.6 trillion parameters, 49 billion active, text-only, 1M-token context, 384K maximum output, MIT licence, and $0.66 per million input against $1.98 output in DeepSeek's off-peak window. DeepSeek V4.1 Pro is a name the company used once, in a release note on 10 September 2026, and has not used since. There is no model card, no price row, no repository and no identifier for it on any surface DeepSeek publishes. So this is not a comparison between two products. It is a comparison between a product and a plan.

What each name actually points at

Start with what is documented, because the two tiers are easy to confuse and the confusion runs in one direction: readers assume the newer number is the bigger model.

DeepSeek V4 Pro — the flagship of the V4 generation, general availability since 13 August 2026 as DeepSeek-V4-Pro-0813, superseding the April preview. A 1.6-trillion-parameter mixture-of-experts with 49 billion active parameters. It is the text-only tier: on DeepSeek's own price sheet the vision row reads "Not supported" for Pro and a check mark for Flash. It carries three thinking-effort levels — low, high and max — with DeepSeek's guidance being low for simple tasks, high for daily agent work, max for harder scenarios. MIT licence, weights published.

DeepSeek V4.1 Pro — a name inside the V4.1 line, which so far has exactly one shipped member. What exists is a sentence from the 10 September release note that shipped DeepSeek V4.1 Flash: that from 04:00 UTC on 14 September, requests to deepseek-v4-pro would route to Flash and be billed at Flash rates, "until V4.1-Pro launches." Developers objected to a production model being swapped underneath them, DeepSeek withdrew the plan and kept serving V4 Pro, and the name survived the withdrawal with no release date attached.

That is the whole documented record. Everything else you will read about V4.1 Pro — parameter counts, context, modality, price — is the family's figures relabelled. The 10 September technology report, a 51-page document on KV-cache compression, does not contain the strings "V4.1-Pro" or "V4.1 Pro" anywhere, and its comparison tables use V4 Pro as a baseline the new model beats, not as a tier it belongs to.

The specification gap, with the sources named

The only figures that can be attributed to DeepSeek V4.1 Pro are inherited, and it is worth being precise about which document each one comes from:

• Backbone — V4 Pro is 1.6T total with 49B active; V4.1 Flash is 552B with 8B active during prefill and 16B during decode, on the new Causal Encoder-Decoder architecture. Both per their Hugging Face cards. No figure exists for V4.1 Pro.

• Context and output — 1M tokens of context and a 384K maximum output, listed identically for both served tiers on DeepSeek's Models & Pricing page. Not restated for V4.1 Pro.

• Modality — V4.1 Flash takes text and images and returns text; V4 Pro is text-only. Whether a V4.1 Pro would close that gap is unpublished, and it is not a safe assumption in either direction: the V4 generation deliberately ran the tiers on different shapes, with Pro the text-only one.

• Licence — MIT across everything that has shipped: DeepSeek-V4-Pro, DeepSeek-V4-Flash, DeepSeek-V4-Flash-0731 and DeepSeek-V4.1-Flash all carry it on their cards. Nothing has been published for V4.1 Pro.

• Benchmarks — none for V4.1 Pro. Every score attached to the name in circulation is a Flash score or a V4 Pro score.

One timing fact is measurable, and it is the reason to stop treating the release as imminent: in the V4 generation, the Pro tier trailed the Flash tier by about two weeks. In V4.1, the Pro name has trailed Flash by more than a month.

The price columns that do exist

Both served tiers are priced on DeepSeek's own sheet, and the shape of the pricing is the most useful thing on it. Prices are off-peak, and they double during two weekday windows — 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays.

• DeepSeek V4.1 Flash — $0.15 per million input and $0.60 per million output off-peak; $0.30 and $1.20 in peak windows.

• DeepSeek V4 Pro — $0.66 input and $1.98 output off-peak; $1.32 and $3.96 in peak windows.

Pro is roughly four and a half times Flash per token and carries a fifth of the concurrency allowance. No successor price has been hinted at for V4.1 Pro, which is itself a signal: a tier with no card has no card price, and anyone quoting one is extrapolating.

Which one to call, and why the question has a real answer

If the reason you are reading this is that you want the strongest DeepSeek model you can put a production path behind, the answer is DeepSeek V4 Pro, and it is one request away. On OrcaRouter it is served at DeepSeek's list price — $0.66 per million input and $1.98 per million output off-peak, passed through with no markup — through the same key and the same OpenAI-compatible endpoint as everything else in a 224-model catalogue. Its 1M context and 384K output ceiling are the same numbers the V4.1 line advertises, and its thinking-effort levels are the ones DeepSeek documents.

If what you want is the new architecture — native image input, the encoder-decoder split, the reduced KV-cache cost the 51-page report is about — then the only released model with any of that is DeepSeek V4.1 Flash. It now serves the deepseek-flash identifier as DeepSeek-V4.1-Flash-1017, the October checkpoint, at $0.15 and $0.60 off-peak on OrcaRouter, which is a quarter of Pro's input price and a third of its output price.

The reason the choice is not merely academic is that the two served models answer different questions. Pro is the ceiling on capability; Flash is the ceiling on throughput per dollar, and DeepSeek's own framing says Flash comprehensively surpasses the previous flagship on performance, cost, speed and total runtime. That is the company's claim about its own Pro tier, and it is the single loudest argument that a V4.1 Pro, when it arrives, will not simply be "V4 Pro but newer."

What would falsify this page

Three things, each checkable in one page load, and any one of them would move V4.1 Pro from speculation to product:

• A changelog entry on DeepSeek's API documentation naming it — the place every V4 release has been announced. The newest entry there is still 10 September.

• A repository in the deepseek-ai organisation with V4.1-Pro in the name. The newest is still DeepSeek-V4.1-Flash.

• A third column on the Models & Pricing page, which today lists exactly two models and is where the vendor states context, modality, features, concurrency and rates.

Until one of those moves, the comparison stays what it is today: one flagship with a card and a price, one name with a release note and a withdrawal, and a family whose newer tier is cheaper and faster than the older one — which is the part readers keep getting backwards.