DeepSeek V4 Pro pricing and specs

Input costs $0.435 per 1M tokens (36.76 ₽) and output costs $0.87 per 1M (73.52 ₽). DeepSeek bills the prompt and the completion separately, so the price of a call depends on how long both are. Ruble figures use the Bank of Russia rate of 84.51 ₽ per dollar for September 2026.

List prices are quoted per 1M tokens, so 1,000 tokens cost a thousandth of that: divide $0.435 for input and $0.87 for output by 1,000. The same rates in rubles are 36.76 ₽ and 73.52 ₽ per 1M tokens. Both prompt tokens and completion tokens are billed.

The DeepSeek V4 Pro context window is 1M tokens (1,000,000). Everything shares that budget: the system prompt, the dialogue history, attached files and the answer the model writes. The figure comes from the DeepSeek model card as of September 2026.

The context window is 1M tokens (1,000,000) and a single response is capped at 384K tokens, so longer output has to be generated in parts. Rate limits are set by DeepSeek per account and depend on your plan rather than on the model, so they are not listed here.

Providers quote their tariffs in US dollars, so the ruble amounts here are a conversion at the official Bank of Russia rate of 84.51 ₽ per dollar. For DeepSeek V4 Pro that gives 36.76 ₽ per 1M input tokens and 73.52 ₽ per 1M output tokens. When the rate moves, the ruble price moves with it.

DeepSeek V4 Pro belongs to the DeepSeek reasoning models family at DeepSeek. Models in the DeepSeek line differ in context size, price per 1M tokens and supported capabilities, so compare them by the numbers on their cards rather than by the name. Every DeepSeek model is listed on the provider page.

Yes, DeepSeek V4 Pro declares a reasoning mode: it works through intermediate steps before it answers. Those steps add output tokens, so a call with a long reasoning chain costs more and takes longer than a plain completion. Capabilities declared for the model: reasoning, tool calling, structured output.

The declared capabilities of DeepSeek V4 Pro are: reasoning, tool calling, structured output. It accepts text as input, and every attachment consumes tokens from the shared context window of 1M tokens. The list comes from the DeepSeek model card and is refreshed with the catalog as of September 2026.

Yes, DeepSeek V4 Pro ships with open weights, so it can be downloaded and served on your own hardware instead of being used only through the DeepSeek API. What you may do with it is set by the license shown on the card. Catalog prices cover hosted API access.