LLM Prices Keep Falling: Should Businesses Celebrate or Be Careful?

On May 21, 2024, Alibaba Cloud cut prices on nine of its large models. The headline: Qwen-Long's input price fell from ¥0.02 to ¥0.0005 per thousand tokens — a 97% cut, about ¥0.5 per million tokens — with output down 90% to ¥0.002 per thousand. Chinese media declared LLM pricing had entered the "li era," prices now measured in thousandths of a yuan. ByteDance cut its model prices the same month. For reference, GPT-4's input price at the time was roughly ¥0.22 per thousand tokens, and Baidu's Ernie 4.0 about ¥0.12.

The price war did not stop in 2024. In January 2025, DeepSeek-R1 listed reasoning benchmarked against OpenAI's o1 at ¥16 per million output tokens, with cache-hit input at ¥1 per million. Today, in early 2026, "models are too expensive" has nearly disappeared as a reason to hold off on AI. The real question has changed: with unit prices falling and falling, should a business celebrate or be careful? Our answer is both — for reasons worth keeping separate.

Where the cuts come from: efficiency, not subsidy

Whether low prices last depends on where they come from, and this round rests mainly on four things. Architectural efficiency: mixture-of-experts (MoE) designs activate only a fraction of a model's parameters per pass — DeepSeek-R1 carries 671B parameters but activates about 37B at a time, decoupling compute cost from parameter count. Inference optimisation: the same hardware serves more requests. Billing design: pricing like "¥1 on a cache hit, ¥4 on a miss" puts an explicit price tag on redundant content. Competition: open-source models set a floor price for capability, forcing closed vendors to justify every premium. Taken together, this is not a burn-money-for-market-share subsidy but a genuine drop in cost structure — low prices are most likely the new normal, not a window.

The case for celebrating: trying things is nearly free

The most tangible dividend is that "let's just try it" no longer needs a budget line. Counting roughly one Chinese character per token, reading a 500,000-character archive once at Qwen-Long's post-cut input price costs about ¥0.25; at the old ¥0.02 per thousand tokens it was around ¥10, and at GPT-4's then price well over ¥100. Scenarios that never used to pencil out — bulk long-document processing, analysing the full history of support tickets, generating content variants at volume — now sit inside "run it once and see" territory. For smaller companies on tight budgets, the entry bar genuinely dropped. If token billing still feels abstract, start with what LLM tokens are.

The case for care: unit price is one corner of the bill

The reasons for caution are just as concrete. First, usage balloons: cut the unit price 97% and multiply call volume by fifty, and the bill grows — cheapness invites "throw everything at the model," which is exactly where waste begins. Second, expensive models on cheap work: reasoning models exist for complex tasks; running format conversion and simple classification through one is embroidery with an excavator, and teams that never tier their models end up paying reasoning prices for non-reasoning work. Third, resending long context: shipping the whole document with every request forfeits the cache discount entirely — everything bills at the miss rate. Stack the three and the savings evaporate; the governance playbook is laid out in how enterprises control AI costs.

What got cheap is the unit price. Total cost is decided by usage governance.

The bigger shift: costs move from the API to people and process

Zoom out, and what this round of cuts really changed is the cost structure of enterprise AI projects. The token bill shrinks as a share of total spend while three items grow: data preparation — shapeless documents yield nothing, however cheap the model; review and acceptance — someone has to check AI output, a standing human cost; operations — prompt upkeep, process adjustment, staff training. Hence a counterintuitive fact: a 97% model price cut does not cut an AI project's budget by 97% — the money migrates from the model side to people and process. Plan budgets around that structure, not around unit prices.

Three moves for businesses

• Tier models by scenario: budget tiers for simple tasks, premium tiers only where reasoning earns its price, private deployment where data is sensitive; open-source models supply the room to choose — see what open-source models mean for enterprise AI costs;
• Stand up a usage dashboard: token consumption and per-task cost by business line and scenario, reviewed monthly, so "why did the bill grow" always has an answer;
• Treat cheapness as a licence to test more, not to care less: the right use of low prices is validating more scenarios at low cost — then keep what pays for itself and switch off what does not.

Back to the question in the title

Celebrate or be careful? Celebrate, because the entry barrier genuinely fell. Be careful, because savings never happen by themselves. What falling prices hand a business is not a discount card but a larger laboratory — used with discipline it is a dividend; used without, it merely discounts the price of waste.