Chinese AI company DeepSeek has quietly released the API version of DeepSeek-V4-Flash, with a focus on improving its agent capabilities.
The main upgrade is giving the model stronger abilities to use tools, operate software environments and complete multi-step tasks.
DeepSeek’s latest model approaches the level of leading frontier models on several agent and coding benchmarks, while keeping inference costs dramatically lower. V4-Flash is priced at just ¥1 per million input tokens and ¥2 per million output tokens. For repeated agent workflows using cached context, the cost drops to only ¥0.02 per million input tokens (around $0.003), making long-running AI agent tasks far cheaper to operate.
For developers, the bigger change is ecosystem integration. V4-Flash now supports OpenAI-compatible Responses API and can work with the Codex ecosystem, allowing developers to call the model through Codex CLI, ChatGPT desktop and VS Code plugins.
Currently, only V4-Flash supports Responses API integration. V4-Pro users will have to wait until early August for similar support.
There is one important limitation: this update is currently available only through the V4-Flash API. The DeepSeek app, web interface, and V4-Pro API experience remain unchanged.