DeepSeek launches public beta API for V4 Flash model

China's DeepSeek on Friday rolled out a public beta application programming interface for its V4 Flash model, touting enhanced agentic capabilities that it says far exceed those of its more powerful V4-Pro-Preview system. The move marks a transition from the model's earlier preview phase to broader public access, intensifying the pricing war between Chinese and American AI labs.bloomberg

Undercutting Western Rivals

DeepSeek V4 Flash, a 284-billion-parameter mixture-of-experts model with 13 billion active parameters, charges $0.14 per million input tokens and $0.28 per million output tokens. That pricing makes it dramatically cheaper than OpenAI's NVIDIA Corporation ↘ 1.47% GPU-powered offerings — even after OpenAI slashed costs on its GPT-5.6 Luna model by 80 percent on July 30, bringing Luna's input price to $0.20 per million tokens.finance.biggo+3

The public beta announcement came with a specific claim about agent performance. "Significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview," DeepSeek said in a post on its website, according to Bloomberg. The model first entered preview on April 24 with MIT-licensed open weights and a one-million-token context window.api-docs.deepseek+2

A Pricing War Reaches a New Phase

The timing is notable. Just one day before DeepSeek's announcement, OpenAI restructured its own pricing to compete at the low end, cutting Luna to $0.20 input and $1.20 output while leaving its flagship Sol model unchanged at $5.00 and $30.00. Yet even after those cuts, DeepSeek V4 Flash undercuts Luna on both input and output pricing.xenospectrum

Data cited by CNBC shows Chinese models can be up to nine times cheaper on a per-token basis than American counterparts, with the share of tokens consumed by U.S. companies on Chinese models staying above 30 percent every week since February. Ion Stoica, a computer science professor at UC Berkeley and co-founder of Databricks, noted that the gap between Chinese open-source models and frontier systems has narrowed from six-to-nine months to just two or three.finance.biggo

What Comes Next

The public beta positions DeepSeek to capture enterprise workloads that prioritize cost over brand loyalty, particularly for high-volume classification, extraction, and coding tasks where V4 Flash's speed advantage matters most. With OpenAI and Anthropic both having filed confidentially for public listings in June, the pressure to defend margins while matching Chinese pricing will only grow as both companies approach their IPO windows.finance.biggo