DeepSeek V4 Pro has officially launched today, bringing major upgrades to agent workflows and developer tools. The new release from the Chinese artificial intelligence startup transitions the model out of its preview phase. This general availability version, designated DeepSeek-V4-Pro-0813, introduces a price adjustment scheduled for August 16.
What is DeepSeek V4 Pro?
The new DeepSeek V4 Pro is a large-scale Mixture-of-Experts model featuring 1.6 trillion total parameters and 49 billion activated parameters. It supports an expansive 1 million token context window, allowing it to process massive datasets. According to the official DeepSeek documentation, the model requires only 27% of single-token inference FLOPs compared to previous versions, making it highly efficient for enterprise deployment.
Agent Capabilities Take Center Stage
The update focuses heavily on autonomous agent capabilities, where systems execute multi-step workflows without human intervention. Benchmark results show massive performance gains. For instance, the model scored 87.9 on Terminal Bench 2.1, up from 72.1 in the preview version. It also achieved a score of 62.7 on DeepSWE, proving its strength in handling complex software engineering tasks.
Flexible Reasoning Effort Settings
Developers can now adjust the thinking intensity of the model using a three-tier reasoning effort parameter. This setting is available for both the flagship model and the faster V4-Flash variant. Users can select low effort for simple tasks, high effort for daily agent workflows, and max effort for complex mathematical or coding problems. This flexibility helps manage operational costs in the growing digital economy.
OpenAI Responses API Support
The API now supports the OpenAI Responses format out of the box, allowing developers to switch providers with minimal code changes. It also integrates with Codex, an open-source framework designed to run agentic tasks. According to reports from The Decoder, this integration simplifies setup for software teams building custom applications.
The model is available today via “Expert Mode” on the official web platform and mobile application. While API identifiers remain unchanged, users should prepare for the upcoming price adjustment on August 16, which will introduce peak and off-peak billing rates to optimize server capacity.
