The general-availability build of DeepSeek-V4-Pro, rolled out across DeepSeek's app, web client, and API on August 13, 2026. It supersedes the April V4-Pro preview while retaining the deepseek-v4-pro API identifier. The API exposes a 1M-token context window, low/high/max thinking-effort controls, and native support for the OpenAI Responses API, including a Codex-specific integration.

DeepSeek reports substantially stronger agent performance than the preview: 87.9 on Terminal-Bench 2.1, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, 61.5 on NL2Repo, and 60.0 on Humanity's Last Exam with tools (42.7 without). Other reported results include CyberGym 83.3, DSBench-FullStack 71.1, and DSBench-Hard 67.2. DeepSeek released the GA checkpoint's weights on Hugging Face under the MIT License; for high/max reasoning effort it recommends allowing up to 384K output tokens.

Model Details

Architecture MOE
Context window 1,048,576
License MIT
Base model deepseek-v4

Benchmark Scores

Benchmark Score Mode
Terminal-Bench 2.1 87.9
DeepSWE 62.7
Toolathlon-Verified 74.1
NL2Repo 61.5
Humanity's Last Exam 60.0 with tools

Paper

frontiermoereasoningcodingagentic

Related