Chinese AI lab DeepSeek has released the official DeepSeek-V4-Flash API in public beta. To use the latest version, callers set the model name to deepseek-v4-flash.
The release comes hours after OpenAI launched a price war by discounting its GPT-5.6 Luna. OpenAI cut input token prices from $1 to $0.2 per 1 million, and output tokens from $6 to $1.20 per 1 million, claiming the steep discount was possible after extracting additional architectural efficiencies.
OpenAI reduced GPT-5.6 Luna input prices by up to 80 percent.
DeepSeek priced V4 Flash 0731 at $0.14 per 1 million input tokens and $0.28 per 1 million output tokens. The model has 284 billion parameters and offers performance similar to Anthropic's Opus 4.8, which is widely believed to span multi-trillion parameters.
For Code Agent tasks in public benchmarks, DeepSeek tested V4-Flash using the DeepSeek Harness minimal mode with max effort level, topp=0.95, and temperature=1.0. The company also uses internal test sets DSBench-FullStack for full-stack development and DSBench-Hard for hard coding agent problems.
The DeepSeek API now supports V4-Pro and V4-Flash via both the OpenAI ChatCompletions interface and the Anthropic interface. The base_url remains unchanged; callers set the model parameter to deepseek-v4-pro or deepseek-v4-flash.
DeepSeek announced that legacy API model names deepseek-chat and deepseek-reasoner will be discontinued in three months, on 2026-07-24. During the current period, these names point to the non-thinking and thinking modes of deepseek-v4-flash, respectively.
In related moves, DeepSeek upgraded both deepseek-chat and deepseek-reasoner to DeepSeek-V3.2. A temporary endpoint serves DeepSeek-V3.2-Speciale at base_url='https://api.deepseek.com/v3.2_speciale_expires_on_20251215', with the same pricing as V3.2, no tool calls, and availability until Dec 15th, 2025, 15:59 UTC.
Earlier, both models were upgraded to DeepSeek-V3.2-Exp and then to DeepSeek-V3.1-Terminus, which maintained original capabilities while addressing user-reported issues. DeepSeek also upgraded deepseek-chat to DeepSeek-V3-0324 and deepseek-reasoner to DeepSeek-R1-0528, and previously to DeepSeek-V2.5-1210, which optimized file upload and webpage summarization.
The company noted that DeepSeek V2 Chat and DeepSeek Coder V2 were merged into DeepSeek V2.5, accessible via deepseek-coder or deepseek-chat for backward compatibility. V2.5 improved general and code abilities, and the API adopted hard disk caching to reduce prices further. Earlier benchmarks showed DeepSeek-V2-0628's win rate against GPT-4-0314 in Arena-Hard rose from 41.6% to 68.3%, and IFEval accuracy jumped from 63.9% to 77.6% for V2-0517.