DeepSeek's V4 Flash Price Shock Continues: 7-Fen Games and Platforms Racing to Subsidize

DeepSeek's V4 Flash price shock is still rippling through the developer world. According to a report from QbitAI on August 4, developers are using the model to build absurdly cheap AI applications: a space shooter game built with V4 Flash costs about 7 fen (0.07 yuan), and a Counter-Strike session costs about 5 mao (0.5 yuan).

Even more striking, overseas platforms are piling on subsidies. OpenCode said DeepSeek Flash consumed 8 trillion tokens on its platform on August 1 alone; Nous Portal offered V4 Flash at 90 percent off for seven days, making it more than a thousand times cheaper than Fable 5 at that discount; and open-source coding agent Cline, which had announced free quota with V4-Flash last Saturday, tripled the free quota after seeing its bill.

The turning point came with V4-Flash-0731, released on July 31. It uses the exact same architecture and model size as the V4 Preview that shipped on April 24 — DeepSeek only re-ran post-training — yet performance jumped dramatically: DeepSWE climbed from 7.3 to 54.4, Cybergym hit 76.7, Terminal Bench 2.1 reached 82.7, and the Artificial Analysis intelligence index rose 10 points to 50, surpassing its own V4 Pro and approaching Opus 4.8.

Price remains DeepSeek's most formidable weapon. A research firm has vouched that DeepSeek is the cheapest known frontier model: US$0.14 per million input tokens and US$0.28 per million output tokens on cache misses.

When a model trails the top players by a wide margin, low price is merely a value proposition. But when the gap narrows to a few points, the casualties of this benchmark race may include more than just the models ranked behind DeepSeek.

The bigger significance of falling costs is ecosystem-level: developers' default model of choice is quietly shifting, more developers are willing to embrace open-source models, and more newcomers can afford to experiment with AI, making product prototypes that once seemed out of reach suddenly buildable.

The next thing to watch is whether DeepSeek completes its toolchain. The article notes that DeepSeek Harness lead Cui Tianyi has posted on X recruiting internal testers for open-source Agent Harness projects, suggesting DeepSeek Code may be on the way — a strong model, rock-bottom prices, and now a coding toolchain being filled in.