On August 17, DeepSeek raised V4 Pro's peak-hour cache-hit price from ¥0.025 to ¥0.30—a 12x jump. The cheapest Chinese frontier model is no longer cheap. The same week, the company fired two more shots: the V4 Pro general release on August 13, closing gaps in Agent capability, and Harness, its in-house Agent framework, on August 14. Price hikes, new products, capability catch-up—three moves pointing to the same judgment: China's LLM competition is pivoting from a price war to a value war.
What this is
DeepSeek's three moves this week are worth unpacking.
V4 Pro general release: Compared with the preview, it closes gaps in Agent capability (AI that autonomously decomposes tasks and calls tools), tool-chain execution (stringing multiple tools together to complete tasks step by step), and software engineering tasks. It also natively supports OpenAI's Responses API (the new-generation interface standard purpose-built for Agents), enabling direct connection to Codex without protocol conversion.
Harness: DeepSeek's in-house Agent runtime framework. It is not a new model but a "shell"—wrapping the model into an application that can get work done. Using LangChain's formula: Agent = Model + Harness. Claude Code and Codex are the same category of product.
Price hike: V4 Pro's cache-hit price jumped 12x, output price up 4.5x. DeepSeek's once-core selling point—"same work, a fraction of the price"—has been significantly weakened.
Industry view
Our editorial judgment: this is not just one vendor repricing; it is an inflection point in China's LLM competitive logic.
On one hand, the past year of China's LLM playbook was "cheapest wins the developers," with DeepSeek as the biggest beneficiary. That low-price model is no longer sustainable—compute supply is tight, and competitors (GLM-5.3, Qwen 3.8 Max) have caught up or surpassed on capability, so the "price" half of price-performance no longer stands out. On the other hand, shipping Harness is also ecosystem catch-up: DeepSeek previously had only a model, no Agent framework, effectively ceding the application layer to others. This move is about building a closed loop of "model + Agent framework," standing in the same tier as OpenAI and Anthropic.
Counterarguments deserve a hearing. A developer who has long tracked API pricing told us: peak-hour price hikes with off-peak unchanged—i.e., time-of-day tiered pricing—amount to more refined "price discrimination," steering latency-insensitive enterprise users to nighttime and passing cost pressure on to customers with rigid needs. DeepSeek may not have truly "stopped being cheap"; it has just gotten better at pricing.
Impact on regular people
For enterprise IT: Projects billed by API call volume will see noticeably fatter invoices this month. We recommend scheduling non-real-time tasks (batch data processing, report generation) to off-peak hours where possible to save money.
For working professionals: Web and app pricing is not directly affected for now, but the developer threshold for subscribing to API services is rising. The strongest signal ordinary users will feel is the meta-fact: "Chinese AI no longer wins on price alone."
For the consumer market: If DeepSeek passes cost pressure to C-end (consumer-facing) free apps, expect trimmed features or stricter usage caps. More evenly matched competition is good for consumers—a single dominant player is what should truly raise alarms.