1. Re-Igniting the Global AI Price War

Chinese artificial intelligence innovator DeepSeek has officially launched the public beta of its latest production model, 'V4-Flash' (Build 0731), initiating a new chapter in the global price war among generative AI providers. According to an extensive operational report published by Reuters on August 3, V4-Flash has established itself as the single most cost-effective commercial foundation model currently operating in the global market.

Data compiled by AI benchmarker Artificial Analysis shows that V4-Flash operates at an estimated average cost of just 3 cents ($0.03 or approximately 43 South Korean Won) per standardized benchmark task. This dramatic price drop places severe margin pressure on Western market leaders:

1/105th the cost of Anthropic’s flagship Claude Fable 5 ($3.15 per task).
1/62nd the cost of OpenAI’s GPT-5.6 Sol ($1.86 per task).
1/29th the cost of domestic competitor Moonshot AI’s Kimi K3 ($0.86 per task).

API unit pricing for V4-Flash has been set at $0.14 per 1 million input tokens and $0.28 per 1 million output tokens. Furthermore, enterprise users leveraging prompt caching receive up to a 98% discount on repeated input sequences, bringing high-volume context processing costs down to near-negligible levels.

2. Performance: Matching Gemini 3.6 Flash

While positioning V4-Flash as a hyper-budget model, DeepSeek has achieved competitive parity with mid-tier industry benchmarks. In the comprehensive "Intelligence Index"—a composite evaluation covering 9 core disciplines including software engineering, formal mathematical reasoning, and automated business workflows—V4-Flash scored 50 points.

This score places V4-Flash on exact parity with Google’s efficiency flagship, Gemini 3.6 Flash, representing a 10-point gain compared to DeepSeek's initial preview release in April. Although it still trails high-tier Chinese rival Kimi K3 (57 points) and flagship frontier models like Claude Opus 5, Claude Fable 5, and GPT-5.6 Sol by 9 or more points, V4-Flash offers an unprecedented performance-to-price ratio for high-volume enterprise tasks.

3. Sparse MoE Architecture and Hidden Inference Costs

Under the hood, V4-Flash employs a Sparse Mixture-of-Experts (MoE) architecture totaling 284 billion parameters, of which only 13 billion active parameters are engaged during any single inference step. By maintaining a 1 million token context window and focusing engineering efforts on post-training optimizations (RLHF and target alignment) rather than raw parameter inflation, DeepSeek drastically minimized compute overhead.

Industry analysts emphasize, however, that published unit rates do not always reflect total expenditure. Models that undergo long reasoning loops or generate verbose step-by-step solutions can accumulate higher costs over complex workflows. During the Intelligence Index evaluation, V4-Flash consumed approximately 206 million output tokens. While this marks a 12% reduction in token generation volume compared to the previous release, enterprise developers must still evaluate total token throughput alongside list prices when calculating overall operational expenses.

4. Corporate Restructuring and Strategic IPO Ambitions

DeepSeek’s aggressive commercial pricing is closely tied to its long-term corporate finance roadmap. As reported by Bloomberg on July 14, DeepSeek is currently finalizing its financial auditing and accounting structure with plans to complete all filings by the end of December 2026. The startup plans to submit an initial public offering (IPO) application on the Chinese mainland exchange in late 2026 or early 2027, eyeing an official stock market debut in 2027.

After causing the original "DeepSeek Shock" last year, the company faced fierce competition from heavily funded Chinese tech giants and well-capitalized startups, including Moonshot AI, MiniMax, Zhipu AI, ByteDance, and Alibaba. The release of V4-Flash is viewed as a decisive move to reclaim market share, build enterprise API traffic, and establish a strong valuation narrative ahead of its public market listing.

Meanwhile, DeepSeek confirmed that its primary flagship model—the 1.6-trillion-parameter 'V4-Pro'—remains in preview and will be released to the general public as soon as final safety and stability optimizations are complete.

[Copyright (c) Global Economic Times. All Rights Reserved.]