For a while, opening the cloud-service console, the screen full of price-cut banners genuinely dizzying. Alibaba Cloud's Qwen-Long input price dropped from 0.02 yuan per 1K tokens to 0.0005 yuan-down 97%. One yuan buys the character volume of five Xinhua Dictionaries-a comparison I double-checked. ByteDance's Doubao top-tier inference input is priced at 0.0008 yuan per 1K tokens, almost scraping the floor. The joy of wool-gathering is back, but this has no winner.
1. Price Blitzkrieg: Hike in the Morning, Cut in the Afternoon
This round of cuts isn't market regulation-it's more like a valuation-driven squid game. DeepSeek is hiking, Anthropic's valuation hit 1.2 trillion, while OpenAI and Google are forced into close-quarters cuts. Someone on Xiaohongshu complained "bought it half a month ago, asked again and it dropped again." This backstabbed anger reflects the model vendors' prisoner's dilemma: performance gaps can't widen, so they can only stomp each other.
360 Zhi Nao said something honest when answering "how to view the price cuts": it accelerates adoption, but may also make enterprises and developers over-rely on cloud vendors, losing autonomy. My reading: every cent of low price you enjoy now may become a chain binding you in the future.
2. Why Is No One Actually Making Money?
The answer hides in the cost structure. Training a hundred-billion-parameter model burns tens of millions of dollars in compute alone; inference cost is a bottomless pit. Algorithm optimization and hardware gains lower costs slowly, but nowhere near "cut 97% and still profit." Baidu's ERNIE Bot attributed cuts to "tech lowering training costs, vendors chasing market share"-that sounds like official spin. The core logic: they're not selling models, they're selling ecosystem positions.
More deadly, open-source models like Llama 4 directly blew up pricing power. When Meta open-sources top models for free, the commercial-API moat shrinks to just "convenience." Users have zero loyalty, always hunting the next value king. Yesterday's Grok 4.5 was king; tomorrow it might be some open-source quantized Qwen3.
3. The Giants' "Free" Open Scheme: Selling Cloud, Not Models
Why can ByteDance, Google, Alibaba slash models to cabbage prices? Because they don't make money from models at all. Models are cloud-service loss leaders; real profit comes from compute, storage, databases, security suites. Hooking into their API at rock-bottom prices means binding your app architecture and data pipelines to their cloud. Once your business runs, want to migrate? Costs are so high you'll give up.
Supermarkets use 50-cent eggs to draw in seniors, then earn it back from the rest of the cart-exact same logic. Independent model companies have no such backend and survive on funding. The end is either acquisition or becoming a plugin in a giant's ecosystem.
4. How Regular People Survive: Don't Pay for "Permanent Low Prices"
Facing this chaos, a few practical suggestions.
First, ditch the "permanent low price" fantasy. Any service promising permanent low prices may signal an unhealthy business model. Today's shorn wool may become tomorrow's service degradation or data leak.
Second, build your own model toolbox. Don't bind to any single model. Use DeepSeek-Coder V2 for code, Gemini Flash for translation, Qwen-Long for long text. Dynamically evaluate value, not die-harden on one brand. You can also use open-source tools like Ollama to locally deploy quantized models and run inference offline, pushing cost near zero. Even if cloud prices hike, you have a fallback.
Third, focus on "value" not "price." In core workflows, stability, data privacy, controllability matter far more than saving a few bucks. One production incident from an unstable model can cost more than a year of API savings. Put critical tasks on proven models, even if slightly pricier.
Fourth, build irreplaceable "AI leadership." Tools change, prices change, but the ability to define problems, decompose tasks, and design workflows is always your core asset. Learn to squeeze model limits with Prompt Engineering, build your own apps with RAG and Agent architectures-these are things you don't need to worry about "price cuts" for.
5. Endgame: AI Becomes Water and Electricity, but the Pipes Belong to Giants
After this price war, the market will leave 3 to 5 base-model giants, plus countless vertical apps built on them. AI is no longer a luxury but infrastructure like water and electricity. We stand at this turning point: the cost of acquiring intelligence approaches zero, but the pipes delivering it are firmly controlled. Do you just want to be a happy wool-gathering user, or the person who controls their own pipes? That depends on how you build your toolbox now.
References
- Alibaba Cloud Qwen-Long price cut (People's Daily Overseas Edition): http://paper.people.com.cn/rmrbhwb/html/2024-06/13/content_26063249.htm
- 360 Zhi Nao on price cuts (Xinhua): http://www.xinhuanet.com/fortune/20240613/ee2a923cc9c44514b46fc4aa5b85a288/c.html
- ERNIE Bot price-cut rationale (Xinhua): http://www.xinhuanet.com/politics/20240613/44aa92732e474f389510eb0954599f10/c.html
- Ollama local quantized model deployment (Zhihu): https://zhuanlan.zhihu.com/p/1993642974563812198
- Prompt Engineering / RAG / Agent architecture (SegmentFault): https://segmentfault.com/a/1190000047650736