DeepSeek's 1,100% Price Hike: The End of China's Cheap AI Era?

DeepSeek raised API prices for its V4 models by up to 1,100% while launching its newest model, V4-Pro, on August 13, closing a chapter in the cheap-inference war Chinese labs have waged for the past year. The most consequential AI news out of China this week was therefore a pricing decision, not a funding round.

The new pricing, effective mid-August, introduces peak and off-peak tiers, with increases ranging from 50% to 1,100% depending on model, token type and time of use. V4-Pro debuts at $0.435 per million input tokens and $0.87 per million output tokens, with the company noting that demand is not evenly distributed across its infrastructure.

V4-Pro is a 1.6 trillion-parameter mixture-of-experts model with a one million token context window, upgraded agent capabilities and adjustable reasoning effort. Its published benchmark results show it beating some leading Western models on coding and cybersecurity tests.

After a year of Chinese labs weaponizing cheap inference, the pivot to monetization admits two readings: margin pressure from compute costs, or confidence in model differentiation strong enough to charge for. Both readings cut directly into the unit-economics assumptions of Western labs.

The backdrop is a broad Chinese funding surge. Chinese companies have raised $31.6 billion across 486 equity rounds year to date, and Crunchbase places Asia's second-quarter funding at a multi-year peak led by Chinese AI. DeepSeek itself is reportedly seeking a valuation of about $74 billion in a round paving the way for a potential mainland listing.

When the cheapest player in the market raises prices elevenfold, the question is no longer how much AI costs, but who can keep subsidizing it. The answer will draw the next map of competition between China and the West.

Key terms explained:

API (application programming interface): The channel that lets developers plug an AI model into their own software, billed by how much they use.

Inference: Running a trained model to produce answers — the recurring cost paid on every request, as opposed to the one-off cost of training.

Token: The unit of text a model processes, roughly a short word or part of one; pricing is calculated per million tokens.

Mixture of experts: A design that splits a model into specialised sub-networks, activating only a fraction on each request to cut running costs while keeping total size large.

Parameters: The numerical values a model learns during training; their count is a rough proxy for its size and capability.

Context window: The maximum amount of text a model can read and hold in a single request.

Peak pricing: Charging more during hours of heavy demand and less outside them, to spread load across the infrastructure.