DeepSeek-V4-Flash API Released with High Cost-Efficiency in Speculative Benchmark Report
DeepSeek has launched the public beta of its DeepSeek-V4-Flash API, with benchmark results from Artificial Analysis and Arena.ai showing performance close to speculative models like GPT-5.6 Luna. The model achieves a significantly lower single-task cost, reportedly 60% cheaper than GPT-5.6 Luna. This highlights the aggressive pricing strategies and optimization techniques, such as deep prompt caching discounts, that AI providers use to lower operational costs. However, the benchmark references futuristic models and dates, suggesting it may be a simulated or speculative projection of the AI landscape in 2026. DeepSeek offers an aggressive 98% prompt cache hit discount on its API, compared to the industry standard of 90%. In Arena.ai's Frontend Code Arena, DeepSeek-V4-Flash-High scored 1,586 points, ranking 7th overall and costing $0.14 to $0.28 per million tokens.
## BACKGROUND
Prompt caching is an optimization technique where frequently used parts of a prompt are stored in memory, allowing the LLM to skip reprocessing them and significantly reduce latency and API costs. The Artificial Analysis Intelligence Index is a composite benchmark that aggregates multiple evaluations to measure overall AI capabilities across reasoning, coding, and mathematics.