DeepSeek officially implemented its latest price adjustment on August 17, introducing a peak-valley pricing mechanism that marks a notable shift for the company long regarded as the industry's cost disruptor.

Under the new structure, peak hours are designated as 9:00-12:00 and 14:00-18:00 Beijing time, with all other periods classified as off-peak. For instance, DeepSeek-V4-Pro during off-peak hours is priced at 0.15 yuan per million tokens for cached input, 4.5 yuan for uncached input, and 13.5 yuan for output — with peak-hour rates doubling these figures.

Prior to this move, DeepSeek had built its reputation on aggressive pricing. In late 2024, the company's V3 model achieved top-tier performance at minimal training costs, and its subsequent R1 model priced its output API at just 3% of OpenAI's o1 offering. In May of this year, DeepSeek announced that after a limited-time 25% discount ended on May 31, the V4-Pro API price would be permanently set at one-quarter of its original rate.

Industry analysts suggest this reversal may stem from mounting pressure on computing resources, while also signaling a broader industry shift from price-based competition toward value-driven competition in China's large model sector.

DeepSeek Rolls Out Peak-Valley Pricing Strategy

At midnight on August 17, DeepSeek introduced its new API pricing structure, departing from the industry-standard flat-rate model by implementing time-based pricing tiers.

According to the company's announcement, peak hours run from 9:00 to 12:00 and 14:00 to 18:00 Beijing time. During off-peak periods, DeepSeek-V4-Flash pricing stands at 0.05 yuan per million tokens for cached input, 1.5 yuan for uncached input, and 4.5 yuan for output. Meanwhile, DeepSeek-V4-Pro is priced at 0.15 yuan, 4.5 yuan, and 13.5 yuan for the same respective categories.

Peak-hour rates double these figures. For DeepSeek-V4-Pro specifically, peak pricing reaches 0.3 yuan for cached input, 9 yuan for uncached input, and 27 yuan for output per million tokens. Compared to the previous pricing of 0.025 yuan cached input, 3 yuan uncached input, and 6 yuan output, peak-hour increases reach as high as 1,100%.

DeepSeek stated that the adjustment aims to optimize resource allocation and encourage users to schedule tasks during off-peak periods.

Notably, DeepSeek is not alone in raising prices. Across China's large model landscape, 2026 has brought a collective wave of price increases. Zhipu AI, which holds the title of "first Hong Kong-listed large model company," has raised prices three times this year alone. Moonshot AI, following the release of Kimi K3, has also significantly increased API pricing, with input costs rising over threefold and output costs nearly fourfold.

Morgan Stanley research from August 9 indicates that average API input prices for Chinese large models reached 4.9 yuan per million tokens in Q2 2026, with output at 21.9 yuan — up approximately 48% and 80% respectively from Q1 2025 levels of 3.3 yuan and 12.2 yuan.

However, even after substantial increases, DeepSeek maintains a considerable price advantage over international competitors. According to Guosheng Securities estimates, DeepSeek-V4-Pro's peak-hour output pricing is roughly 1/13 of Claude's flagship model, and only about 1/25 during off-peak hours.

Price Strategy Drives Record-Breaking Usage

Tianyancha records show that DeepSeek's affiliated company, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., was established in 2023. In late 2024, DeepSeek-V3 debuted with industry-leading performance achieved at remarkably low training costs, rivaling top models like GPT-4o.

Shortly thereafter, in January of last year, the company launched DeepSeek-R1, continuing its cost-performance tradition. The model's API pricing was set at 1 yuan per million tokens for cached input, 4 yuan for uncached input, and 16 yuan for output — with output pricing at just 3% of OpenAI's o1.

This extreme cost-effectiveness propelled DeepSeek to rapid prominence. In April of this year, the company released its V4 series preview models, including the 1.6T-parameter DeepSeek-V4-Pro and the 284B-parameter DeepSeek-V4-Flash. On April 26, DeepSeek announced that cached input prices across all API models would drop to one-tenth of launch pricing, with V4-Pro receiving an additional limited-time 25% discount, bringing cached input to just 0.025 yuan per million tokens — a global record low.

The following month, DeepSeek continued its aggressive pricing strategy, announcing that after the limited-time discount concluded on May 31, V4-Pro's API price would permanently settle at one-quarter of the original rate. Post-adjustment, V4-Pro pricing became 0.025 yuan for cached input, 3 yuan for uncached input, and 6 yuan for output per million tokens.

This low-price approach has driven explosive API usage growth. According to OpenRouter data, from July 27 to August 2, DeepSeek-V4-Flash ranked first globally with 7.22 trillion tokens of weekly usage, while V4-Pro placed fourth. Following the official release of V4-Flash on July 31, the model topped usage charts for two consecutive weeks, recording 8.83 trillion and 11.2 trillion tokens respectively.

Low pricing does not mean compromised quality, however. DeepSeek's official benchmark data shows that V4-Pro outperforms Opus-4.8 across multiple test suites including Terminal Bench 2.1, Cybergym, DeepSWE, and AutomationBench (Public). Even against Fable5, V4-Pro holds its own in several categories.

A new social media term, "DeepSeek kill line," has emerged following V4's success, originating from Artificial Analysis's cost-performance scatter plot. The chart plots per-task cost against intelligence scores, with DeepSeek-V4-Flash's position creating a dividing line: models with comparable performance but higher prices, or weaker performance at similar costs, are relegated to the "kill zone," facing potential obsolescence.

Why Would the Price Disruptor Change Course?

While DeepSeek's extreme value proposition has attracted massive user adoption, surging demand has placed significant strain on its infrastructure. On August 4, the open-source AI coding platform OpenCode reported capacity issues with DeepSeek-V4-Flash due to unprecedented traffic, causing errors that required emergency fixes.

Users reported that the official V4 Flash API was nearly unusable during the morning of that day, though DeepSeek later confirmed the issues were resolved and services fully restored.

These server pressures reflect a broader reality of tightening computing resources. Data from the China Development Forum 2026 annual meeting shows that daily token usage in China surpassed 140 trillion in March of this year — a thousand-fold increase from the 100 billion recorded in early 2024. Yet supply-side capacity has not kept pace with this explosive demand.

High-end GPUs and HBM memory remain in critically short supply, with Nvidia's Blackwell and Hopper series unable to meet demand. China Academy of Information and Communications Technology data indicates that domestic AI computing demand grew 417% year-over-year in Q1 2026, while effective supply increased only 128%, widening the supply-demand gap. Moonshot AI's Kimi, for example, suspended new user subscriptions after Kimi K3 went viral, redirecting all available computing resources to existing subscribers.

Analysts suggest DeepSeek's pricing pivot also signals a broader industry transition from price competition toward value competition. Morgan Stanley's research notes that China's large model industry is building a healthier commercial environment, with competitive focus shifting from price to model capability. The firm argues that long-term competitive positioning depends on model intelligence rather than pricing, as advancing frontier capabilities requires substantial training and computing investment. Relying solely on low prices to capture market share would compress margins and limit funding for next-generation research.

Zheshang Securities similarly observes that DeepSeek's price increase challenges the pessimistic expectation of endless price wars in the model industry, raising the valuation benchmark for domestic large model companies. Radar Finance will continue monitoring DeepSeek's ongoing developments.