WAIC 2026: Chinese AI 'price war' turns to 'premium tiering'

If there was one question that echoed across the exhibition halls at this year's World AI Conference (WAIC), it was this, "What should AI actually cost for average users?"

Though most AI services provide free chat options, many users are looking for model API keys that can drive their AI agents to get real jobs done instead of just talking.

The price question hung over nearly every booth. Bai Chuanxu, an AI architect with Shanghai-based AI unicorn MiniMax, defended his company's 40-yuan (about $5.5) monthly subscription as sustainable because MiniMax models are built for cost efficiency. 

"Models keep improving," he said. "GPT-3.5 cost $20 a month back in the day. Today's domestic 40-yuan subscription models are generally much stronger than that."

At iFlytek, product manager Ru Huiya pointed to the company's new agent desktop product, which offers daily free quota for light users – though she did not specify when the free tier would end.

An Alibaba Cloud staffer at the company's exhibition booth compared the moment to mobile network's 3G-to-4G transition. "Plans were expensive at first, then came down," he said. "AI may follow a similar path." 

He added that the 40-yuan tier is unlikely to be sustainable, while the 200-yuan ($27.5) tier has better prospects.

The conference was bookended by a development that sharpened the debate. Two days before WAIC kicked off, Moonshot AI launched Kimi K3 – a 2.8-trillion-parameter flagship model priced at $3 per million input tokens and $15 per million output ($0.3 on cache hit). That makes it roughly 17 times the output cost of DeepSeek V4 Pro. It also topped the LMArena front-end coding leaderboard on launch, signaling a deliberate strategy – not every model has to compete on price.

Several exhibitors told CGTN that AI application paradigms have not settled, making price prediction difficult. At current compute supply and demand levels, many said, the 40-yuan monthly subscription is hard to sustain.

Chinese model providers, riding OpenClaw's popularity early this year, rushed to offer 40-yuan monthly coding plans with 18,000 model requests and no explicit per-request token limits. A few months later, many of those plans have been discontinued or are hard to purchase. Providers are steering users toward per-token billing, with prices generally rising. Meanwhile, many internet services have launched desktop AI agent clients with daily free token allowances. Long-time observers of coding plan communities note that many users find DeepSeek's pay-per-token pricing affordable even without a subscription.

The overall picture emerging from WAIC is market segmentation. Premium models like Kimi K3 carve out a high-performance tier. Budget-conscious users gravitate toward pay-as-you-go. The middle ground – fixed-price subscriptions – appears to be shrinking.

"Is the 40-yuan plan sustainable?" one exhibitor asked. "Not at current cost structures. However, that doesn't mean AI is expensive. It means the market is still figuring out what AI should cost."