Model Watch · DeepSeek

DeepSeek V4

3 models carry the DeepSeek V4 name, 3 of them callable today. Here's what separates them, and what each one costs.

From DeepSeek's own docs · updated 2026-09-05 10:32 UTC

The family started with DeepSeek-V4-Flash on 2026-07-31 and the newest is DeepSeek-V4-Flash-Vision-Exp from 2026-08-21. Running one of these for a month costs anything from $0.77 on DeepSeek-V4-Flash-Vision-Exp to $2.32 on DeepSeek-V4-Pro, for identical work.

Every DeepSeek V4 model

The highlighted row is the newest one DeepSeek has actually released. The monthly figure is 700 words in and 500 out, twenty times a day, at list prices. Change the workload.

Which one should you be using?

DeepSeek-V4-Pro is the newest released model in the family, out on 2026-08-13. Newer is not automatically better for your job, but it is usually the one the lab is optimising, and on this table it is rarely the most expensive.

Why there are so many numbers

A point release is a new model, not a patch. DeepSeek ships a new one under the same family name when the training or the tuning changes, and the old one usually keeps working at its old price until it's deprecated.

That's why two models in this table can hold the same amount of text and cost different amounts. You're picking a specific model, not a family, and the API name in each model's page is what actually goes in your code.

What each one is for

• DeepSeek-V4-Flash-Vision-ExpThe multimodal one, and DeepSeek calls it experimental in its own release note. Priced like Flash.
• DeepSeek-V4-ProA 1.6T mixture of experts on a million-token window for a third of what the US flagships charge. Run it outside 01:00 to 10:00 UTC on a weekday and DeepSeek halves the bill.
• DeepSeek-V4-FlashForty-four cents a million on a million-token window, twenty-two off-peak. The open weights went up the same day.

Keep reading

DeepSeek prices in full · find a model by budget · compare any two