DeepSeek Launches V4 Pro Model with Open Agent Tools and Higher API Rates

DeepSeek released the general availability version of its flagship artificial intelligence model, DeepSeek-V4-Pro, on Thursday, August 13, 2026.

The Chinese technology firm deployed the update, designated as version DeepSeek-V4-Pro-0813, across its web platform, mobile application, and application programming interface (API).

Along with the main model, the company launched DeepSeek Harness v0.1, an open-source framework published on GitHub under the MIT license to build and run software agents.

The new release targets long-running software workflows and complex automation tasks.

DeepSeek built the model on a Mixture-of-Experts architecture containing 1.6 trillion total parameters and 49 billion active parameters per token.

The architecture supports a context window of 1 million tokens and generates up to 384,000 output tokens in a single request.

In its official documentation, DeepSeek described the core functional focus of the release:

"Major Agent upgrades with strong production gains!"

DeepSeek added three adjustable reasoning effort levels to the model API.

Users can set reasoning effort to low for standard queries, high for routine agent execution, or max for complex technical tasks. The API includes native support for the OpenAI Responses API and integration with Codex through a setup interface.

Accessing the flagship model through the company API will require higher payments starting Sunday, August 16, 2026 at 16:00 UTC.

DeepSeek is replacing its existing flat API rates with dynamic peak and off-peak pricing structures.

Off-peak usage offers a 50% price reduction to encourage automated jobs outside high-demand hours.

The company announced the structural billing changes in its release notice:

"With the V4 lineup release, we're updating our API pricing and introducing peak and off-peak rates. Off-peak rates are 50% lower than peak, enabling more flexible workload scheduling."

Data from independent benchmarking firm Artificial Analysis shows the peak list rates for DeepSeek-V4-Pro-0813 at $1.32 per million input tokens and $3.96 per million output tokens.

These figures equal nine times the input cost and 14 times the output cost of the smaller DeepSeek V4 Flash model.

DeepSeek lists V4 Flash at $0.14 per million input tokens and $0.28 per million output tokens.

Official API documentation lists cache hit pricing for V4 Pro at $0.003625 per million tokens, cache miss input at $0.435 per million tokens, and output at $0.87 per million tokens.

The release of DeepSeek Harness v0.1 expands the company's footprint from raw model hosting into developer runtime software.

Software teams often use dedicated agent execution tools like Anthropic's Claude Code to let models edit files, run shell commands, and handle multi-step programming jobs. DeepSeek Harness provides a modular, plugin-based runtime where engineers can swap out file handlers, tool caller loops, and execution environments.

Performance evaluation records published by OpenRouter show DeepSeek-V4-Pro reaching an accuracy score of 80.6% on the SWE-bench Verified coding benchmark in maximum reasoning mode.

The model recorded 93.5% on LiveCodeBench and 90.1% on the GPQA Diamond scientific reasoning test under maximum reasoning settings. On terminal management tasks, the model scored 67.9% on Terminal Bench 2.0.

Developer access to the DeepSeek-V4-Pro-0813 checkpoint remains active immediately through the company's web interface under Expert Mode and via standard API endpoints ahead of the August 16 price transition.