DeepSeek's V4.1 Pro has a name but no model card, price or release date

One flagship, one name

People searching for DeepSeek's best model tend to land on a comparison of V4.1 Pro and V4 Pro. As OrcaRouter reported, that comparison is lopsided. V4 Pro is a released product. V4.1 Pro is a name the company has used only once.

V4 Pro became generally available on August 13, 2026 as DeepSeek-V4-Pro-0813, replacing an April preview. It has a model card (the public spec sheet), a price list and a working endpoint, which is the address developers send requests to.

V4.1 Pro showed up in a release note dated September 10, 2026, and has not been used since. DeepSeek has published no model card, price entry, code repository or model identifier for it on any channel. So this is a comparison between a product and a plan.

The one sentence that exists

The release note announced DeepSeek V4.1 Flash, a cheaper, faster tier. It also said that from September 14 at 04:00 UTC, requests to the deepseek-v4-pro identifier would be routed to Flash and billed at Flash rates until V4.1 Pro was released.

Developers objected to having a production model swapped without their say. DeepSeek withdrew the plan and kept serving V4 Pro. The V4.1 Pro name stayed in circulation, with no release date attached.

The 51-page technical report published the same day is about compressing the KV-cache. That is the memory a model keeps of the text it has already processed, so shrinking it makes long inputs cheaper to handle. It never uses the phrase V4.1 Pro. Its tables treat V4 Pro as the baseline the new model beats, not as a tier the new model belongs to.

What V4 Pro actually is

V4 Pro is a mixture-of-experts model, a design where only part of the network works on each piece of text. It has 1.6 trillion parameters in total, of which 49 billion are active at a time. It handles text only. On DeepSeek's own price table, the vision row for Pro reads "not supported", while Flash has a check mark.

It accepts a 1M-token context, meaning the amount of text it can consider at once, and can produce up to 384K tokens of output. It offers three reasoning-effort levels: low, high and maximum. DeepSeek suggests low for simple tasks, high for everyday agent work and maximum for harder cases. The weights are public under the MIT license, a permissive open-source license.

The V4.1 Pro numbers you may see online are borrowed. Total and active parameters, context, output limit, modalities, license and benchmark scores all lack a V4.1 Pro entry. Any score attached to the name is really a Flash score or a V4 Pro score.

The price list that exists

DeepSeek lists two served tiers, with prices per million tokens. The prices below are off-peak. They double during two weekday windows, UTC Monday to Friday 01:00 to 04:00 and 06:00 to 10:00, except on Chinese public holidays.

• V4.1 Flash: $0.15 input and $0.60 output off-peak; $0.30 and $1.20 at peak.
• V4 Pro: $0.66 input and $1.98 output off-peak; $1.32 and $3.96 at peak.

Per token, Pro costs roughly four and a half times as much as Flash, and it comes with one fifth of Flash's concurrency allowance, the number of requests you can run at once. A tier with no model card has no card price, so anyone quoting a V4.1 Pro price is extrapolating.

Which one to call

For the strongest DeepSeek model you can put into production today, the answer is V4 Pro. The outlet, which sells access to many models through one interface, says it passes DeepSeek's list price through with no markup. Its 1M context, 384K output cap and effort levels match the figures the V4.1 line advertises.

For the new architecture, there is one released option: V4.1 Flash. It accepts text and images, and it uses a new encoder-decoder design with 552 billion total parameters. It activates 8 billion during prefill, when the model reads your prompt, and 16 billion during decoding, when it writes the answer.

The two answer different needs. Pro is the capability ceiling, while Flash is built for the most work per dollar. DeepSeek itself says Flash beats the previous flagship on performance, cost, speed and total run time. That is the best argument that V4.1 Pro, when it arrives, will be more than a refreshed V4 Pro. It also means the newer name may not be the bigger model.

What would settle it

The source names three things you can check in a page load, any of which would turn V4.1 Pro from rumor into product. First, a changelog entry in DeepSeek's API docs, where every V4 release has been announced; the latest entry there is still September 10. Second, a repository with V4.1-Pro in its name in the deepseek-ai organization, where the newest is still DeepSeek-V4.1-Flash. Third, a third column on the Models and Pricing page, which currently lists only two models.

Until one of those changes, the situation stays as described: one flagship with a spec card and a price, one name with a release note and a retraction, and a product line whose newer tier is cheaper and faster than the older one.