DeepSeek V4.1 Pro Release Status and API Availability

DeepSeek has mentioned V4.1 Pro, but has not announced a public release date or listed a public API model ID for it. The deepseek-v4-pro ID currently serves V4-Pro-0813.

Last checked

What has DeepSeek confirmed?

QuestionCurrent statusEvidenceLast checked
Has DeepSeek announced a release date?DeepSeek mentioned V4.1 Pro in the September 10 Flash release. It gave no public launch date.September 10 announcement
Can I use it on the web or in the app?DeepSeek has not confirmed public access or a public beta for V4.1 Pro on its website or in its app.DeepSeek release announcements
Can I call V4.1 Pro through the official API?No V4.1 Pro ID appears in the public model table. deepseek-v4-pro serves DeepSeek-V4-Pro-0813. Check whether your endpoint lists V4.1 Pro.Official API model table
Can I download the weights?DeepSeek’s Hugging Face V4 collection has no model card or downloadable weights listed for V4.1 Pro.DeepSeek’s V4 collection
Has DeepSeek published API prices?DeepSeek has published no V4.1 Pro token rates. Its listed Pro prices apply to V4-Pro-0813.Official API pricing

September 14 was the proposed date for rerouting existing V4 Pro requests to Flash. DeepSeek later confirmed that V4 Pro service would continue. The V4.1 Pro release timeline puts that change alongside the dated Flash releases.

Which model can I use now?

V4 Pro for text input
Set model to deepseek-v4-pro to call DeepSeek-V4-Pro-0813. Use this ID for a text prompt such as a patch review or a question about a function.
V4.1 Flash for text or images
Set model to deepseek-flash to call DeepSeek-V4.1-Flash-1017. To diagnose a visual bug, include a screenshot with the relevant code.

See the V4.1 model and release guide for the version history, or compare V4 Pro and V4.1 Flash for coding using a worked cost example.

Did the November web tests identify V4.1 Pro?

On a private set of 50 agent tasks, an evaluator reported that the DeepSeek web service passed 32 tasks on all three attempts, compared with 24 for the earlier V4 Pro baseline. The evaluator could observe the answers but could not verify which checkpoint served them. See the November LobeBench report and test conditions.