DeepSeek-V4-Flash tops global call-volume ranking again as API prices officially rise

Effective 00:00 on August 17, DeepSeek officially implemented its revised API pricing, adopting peak-valley rates in which off-peak prices are half of peak-hour prices, with V4-Pro seeing increases of up to 1,100% during peak hours. The adjustment arrives as the company's V4-Flash official version tops the global weekly call-volume ranking for a second consecutive week.

The figures come from a report published by National Business Daily on August 17, which estimated from OpenRouter data that global AI model call volume totaled 75.3 trillion tokens last week (August 10-16), up 9.13% week over week. Chinese models accounted for 36.84 trillion tokens, up 7.56%, while U.S. models reached 10.26 trillion tokens, up 11.89%, keeping China's weekly call volume ahead of the U.S. for the 16th consecutive week.

The top two spots globally were both Chinese models: DeepSeek-V4-Flash-0731 (the official V4-Flash release) held first place for a second straight week with 11.2 trillion tokens, up 27%, and Tencent Hy3 ranked second for the second week running at 9.95 trillion tokens, up 24%. The preview build DeepSeek-V4-Flash-0423 placed fourth with 4.83 trillion tokens, down 18%.

The price move landed alongside the call-volume data. Starting August 17, DeepSeek switched to peak-valley pricing, charging half the peak rate during off-peak hours, with peak-hour increases reaching as high as 1,100% for V4-Pro. Calling a leading model during demand peaks now costs noticeably more, while off-peak usage keeps some pricing flexibility.

On the other side of the ranking, OpenAI's GPT-5.6 Luna entered the top three for the first time with 5.31 trillion tokens, up 20% for a third consecutive week of growth. OpenAI had earlier announced optimizations to GPT-5.6 series pricing and performance, including an 80% price cut for GPT-5.6 Luna; Artificial Analysis estimates Luna's average cost per single benchmark task at about $0.05, or one-sixtieth of Anthropic's flagship Claude Fable 5.

The contrast in pricing strategy is striking. On one side, China's leading models are raising prices as capability and call volume grow, while on the other, U.S. vendors keep cutting prices on mid- and low-tier models to court developers — the once-wide price gap between Chinese and American models is narrowing. Morgan Stanley data cited by the report shows Chinese models' average input and output API prices were just 6% and 5% of U.S. peers in Q1 2025, rising to 19% and 14% by Q2 2026.

Also notable: Claude Opus 5 entered the ranking for the first time at No. 8 with 2.67 trillion tokens, up 102%, while NVIDIA's Nemotron 3 Ultra (free) dropped out. The key question now is whether DeepSeek's call volume can keep growing after the price hike, and whether other Chinese vendors follow with similar time-based pricing.