LLMs Are Getting More Expensive — And Customers Are Happy About It

I saw some numbers a few days ago that made me pause.

Zhipu AI raised API prices by 83% in Q1 2026.
Not only did customers not leave — usage surged 400%.

In any traditional industry, doubling your prices and gaining customers is absurd. In the AI large language model space, it’s actually happening.

Last week, Anthropic followed suit: Claude Enterprise dropped its $200 per seat per month package in favor of a $20 base fee plus consumption-based billing. For heavy users, it’s actually more expensive. But Anthropic is betting that customers willing to pay more for compute are precisely the customers they want to keep.

This isn’t simply “price increases.” This is a paradigm shift in how AI companies think about monetization.

Why is this happening?

Layer one: Supply-demand reversal.

Many assume the AI model market is a buyer’s market — more models, lower prices. But the reality is different: capability gaps between top-tier models are narrowing, yet supply of genuinely differentiated models isn’t “oversaturated.” When companies start integrating AI into core business workflows, “which model” matters more than “at what price.” A model that’s 3% less accurate will cost you double the time fixing its mistakes.

Layer two: Customer mix shift.

Early AI adopters were individual developers and hobbyists — highly price-sensitive. 2026’s AI adopters are enterprise customers with completely different decision-making logic. For enterprises, AI API costs are already a small fraction of total costs. What they care about is SLA, reliability, and technical support. “20% cheaper but potentially unstable” is far less appealing than “20% more expensive but I know exactly when it’ll fail.”

Layer three: Differential pricing becomes possible.

Zhipu’s strategy is clever: GLM-5.1 surpassed Claude Opus 4.6 on SWE-bench Pro — the first time a domestic Chinese model achieved this on that benchmark. When the capability gap narrows, pricing power returns. When you can do things others can’t, you earn the right to charge more.

But here’s my concern.

This price hike wave is essentially capturing “AI落地红利” — companies are integrating AI into core operations, and during this window, model providers have pricing power. But this window won’t stay open forever. If Llama 4 open-source models continue improving, if more companies shift to on-premise deployments, the foundation of this pricing power erodes.

My take: these increases are justified in the short term, but sustainability is limited. A new round of price cuts may come in 2027, triggered when inference cost declines outpace capability improvements.

For developers, the key takeaway: don’t bake AI API costs into your business model assumptions. Prices fluctuate — and not always downward.

— Lin Rui, writing from Shenzhen

Your turn: What percentage of your project costs are AI API fees? Would you pay a 20%+ premium for meaningfully better performance?