DeepSeek Ships V4-Pro, Then Raises Prices as the Model Goes Live

DeepSeek formally released its V4-Pro model on 13 August 2026, on its app, website and API. The general availability version, designated DeepSeek-V4-Pro-0813, is built around agent capabilities: tasks where the model uses tools, executes code and completes multi-step workflows without a human checking in at every step.

The model handles a context window of up to 1 million tokens and can produce outputs as long as 384,000 tokens, running in either a thinking or a non-thinking mode depending on how much reasoning a task needs. The API was updated to work with OpenAI's Responses format and includes built-in support for Codex integration, with thinking effort now available at three levels: low, high and max.

The pricing story is the sharper edge of the release. V4-Pro launched at a flat rate of $0.87 per million output tokens, and a price rise to $3.96 per million at peak hours took effect on 16 August, three days after general availability. That is more than a fourfold increase inside a single week, and it lands the model much closer to the pricing of the frontier models it is positioned to compete with, rather than the aggressively cheap rates that first made DeepSeek's models a talking point.

It is a familiar pattern in this market: launch at a price low enough to force a conversation, then adjust once demand and the real cost of serving the model become clear. It does not make the release less significant. A 1 million token context window with genuine agent capabilities, shipped by a lab outside the usual US frontier labs, keeps the pressure on pricing and capability across the board, whatever DeepSeek eventually charges for it.

For teams evaluating models for agent workloads, the practical move is to price a real task, tool calls, retries, a realistic context length, at the new peak rate before committing, not the launch rate that made headlines three days before it changed. A model's opening price in this market is increasingly a marketing number, not a planning one.