DeepSeek has launched the public beta of its new API, DeepSeek-V4-Flash, which has significantly enhanced agent capabilities, achieving a benchmark score of 82.7 on Terminal Bench 2.1. This improvement suggests that the need to use more costly models, like the larger Pro-Preview, is eliminated, aligning with current trends where AI developers prioritize more efficient, cost-effective models that can manage longer agent loops. The new API also offers native support for the Responses API format, facilitating smoother integration into AI agent development environments.

DeepSeek AI: DeepSeek AI develops large language models and AI infrastructure tools focused on accessibility and performance. Through its official account, the company announced the V4-Flash API launch and highlighted upgrades to agent capabilities along with configuration details in its API documentation.
DeepSeek-V4-Flash: DeepSeek-V4-Flash is an AI model API developed by DeepSeek AI for efficient inference and enhanced agentic workflows. It has been released as a public beta with native support for the Responses API format and full adaptation for Codex. The update positions it as a strong alternative for long-running agent tasks by delivering performance that surpasses larger model previews.

API Standards: Native support for formats like Responses API is becoming essential for seamless integration in AI agent development environments.
Model Efficiency: AI developers are prioritizing cost-effective models that handle extended agent loops without relying on larger, more expensive alternatives.