Doubao's new fees spark debate, marking the end of the free era for large models as capital subsidies recede and users pay for frequent use.
May 2025 · Business Observation
Two days ago, the sixth spot on Weibo's trending list carried a topic that made people's hearts skip a beat — #DoubaoNowCharges#.
ByteDance's AI assistant Doubao switched from free to usage-based pricing. The moment the news broke, the user community erupted. "Even Doubao costs money now," "The free lunch for AI is over" — complaints like these flooded the comment sections. To many ordinary users, this felt like a betrayal: wasn't AI supposed to be for everyone? Why the sudden about-face?
But if you zoom out, you'll see a harsher truth: the "free era" of large language models was never the norm. It was a carefully engineered subsidy play, no different from the food-delivery wars or the ride-hailing wars of years past.
Now, the subsidies are receding.
1. It's Not That It "Got More Expensive" — It Was Never Cheap to Begin With
Few people realize how much money burns behind a single conversation with a large language model. A rough but telling figure circulates in the industry: training a GPT-4-class model consumes enough electricity to power a family's air conditioning for several lifetimes. And inference costs — the computing power consumed in the split second when you ask a question and the model answers — are equally staggering.
By industry estimates, an AI assistant with tens of millions of daily active users faces annual inference costs in the hundreds of millions of dollars. When user behavior shifts from "trying it out" to "using it daily at high frequency," companies are no longer dealing with a question of "offering a subsidy" — they're facing "losing the cost of a sports car every day it stays free."
Doubao's pricing is hardly outrageous. At the unit prices disclosed so far, light users — those who ask a few questions and edit a couple of paragraphs each day — would spend about the cost of a cup of milk tea per month. The ones who actually feel the sting are precisely the heaviest users: developers making API calls, creators using AI to draft long-form content, and office workers processing data in bulk with AI. In other words, the price pain isn't felt by those who "can't afford it" — it's felt by those who "can't live without it."
And that's the real business logic of the large-model industry: light users build scale, heavy users generate revenue. Just like every internet product before it.
2. History Doesn't Lie: Every "Free Lunch" Eventually Expires
Looking back, a clear pattern emerges — every "free" offering on the internet eventually finds a way to balance the books.
The most classic case is cloud storage. In the 2010s, Baidu Netdisk, 115 Netdisk, and 360 Cloud Drive launched a fierce "free war," handing out terabytes of space for nothing while users frantically uploaded. And then? Costs became unsustainable. 360 Cloud Drive shut down entirely, 115 pivoted to a membership model, and Baidu Netdisk started throttling speeds — pay up or suffer download speeds of a few dozen KB/s. Users went from outrage to acceptance, then quietly to auto-renewal. Today, Baidu Netdisk's premium membership is one of the most stable revenue streams in the cloud-storage industry. Nobody talks about "free" anymore.
Netflix's story is much the same. Its founder said internally back in the 2000s: "Freemium is the highway to bankruptcy." Netflix never offered a free tier beyond the trial period, choosing instead to win you over with reputation and content before walking you through the paywall. To this day, while every other streaming service bleeds money, Netflix remains profitable.
And ChatGPT? OpenAI's free tier has always existed, but the experience gets more fraught by the day: GPT-4 usage caps, slower response times, peak-hour queues. Want it faster or better? That'll be $20 a month. At its core, this is "neutered free tier + full-featured paid tier" — the same playbook as Baidu Netdisk's throttling, just wearing a different outfit.
The internet's logic of "free" never truly held up. What's called "free" is just a company picking up the tab for you upfront — and then, at the moment you need it most, handing you the bill.
3. Saying Goodbye to Free Isn't "Losing Your Conscience"
The backlash to Doubao's pricing change is, at its emotional core, about one thing: users got used to it.
The past two decades of China's mobile internet can be summed up as a history of "subsidy-driven conditioning." Food-delivery discounts, ride-hailing subsidies, free bike-sharing, buy-one-get-one streaming deals — users' willingness to pay has been trained down to nearly zero. Any move from free to paid triggers a storm of public outcry. But experience tells us that after the storm passes, the people who should renew, do — as long as the product itself is irreplaceable.
Is Doubao worth paying for? The answer varies by person. But the fact that ByteDance dared to take this step now says at least two things: first, market validation is complete — user retention rates prove Doubao is a "hard necessity," not a "toy"; second, industry consensus is forming — AI products can no longer burn investor cash to buy users; they must generate their own revenue.
In fact, since the second half of 2024, the entire large-model industry has been collectively "coming ashore": OpenAI raised API prices, several domestic large-model vendors adjusted their free quotas, and more vertical AI tools are shifting to subscription models. The free strategy went from "user-acquisition weapon" to "operational burden." By 2025, the big players have finally stopped pretending.
4. In the Paid Era, Users Are the Real Counterparty
One last point that many overlook: paying isn't a bad thing — for users, it's actually a good thing.
Free products operate on the logic that "you are the product." When you chat with a free AI, your data is either feeding the model or being monetized on the side. A paid product, by contrast, offers at least one basic thing: a contractual relationship. You pay, you get a service commitment, and if you're not satisfied, you can walk away. In the free era, you didn't even have the right to be "dissatisfied" — because you hadn't paid anything.
Moreover, only with a stable revenue stream does a company have the incentive to keep improving the product, refining the experience, and protecting privacy and security. You might think the free Doubao is fine today, but how do you know it won't shrink its services or downgrade its model quality tomorrow because it can't afford to maintain the team? Paying means sustainable iteration and a guaranteed standard of delivery.
So, Doubao's pricing change doesn't warrant outrage. It's just a mirror.
It reflects a truth: the real path to AI for everyone isn't endless free access — it's fair pricing. It reflects the endpoint of the internet's free philosophy — subsidies always recede eventually. And it reflects each user's choice: are you willing to pay for a good product?
The free era of large language models is over.
But the good era of large language models is just beginning.