Written by: Xiao Bing

On August 7, OpenAI announced that ChatGPT's weekly active users had reached 1 billion, while also making GPT-5.6 Luna available to free users, with no limits on text conversations.

The day before, several DeepSeek API users received notifications: The company plans to raise API service prices overall in the near future, "with a significant increase expected"; specific prices and implementation times will be announced later.

This seems a bit counterintuitive.

For the past two years, Chinese large model manufacturers have excelled in the narrative of extreme cost-effectiveness. DeepSeek previously offered performance close to GPT-4 levels at less than one-thirtieth of OpenAI's price, earning the title “price butcher” in the industry, and Liang Wenfeng became known as “Liang Saint.” The consensus was: The core advantage of domestic large models is to bring the cost of AI use down to rock bottom.

Now the floor is shaky, OpenAI has become the provider giving things away for free to the world, while Chinese manufacturers are starting to talk about charging fees and increasing prices. In this shift between offense and defense, has the position really changed?

OpenAI: Free is not charity

First, let's clarify one thing: GPT-5.6 Luna is not OpenAI's strongest model. It is positioned in the medium capability range, more than enough for handling everyday conversations, simple writing, and basic translations, but for complex reasoning and multi-step code analysis, a more advanced SOL series is still needed. OpenAI is taking a "good enough" model to cover the broadest user scenarios.

This strategy has confidence, but also comes with costs.

The confidence comes from the cost side. Over the past 18 months, the unit token cost curve for large model inference has steeply dropped: model architecture optimization, quantization technology maturity, and upgrades to inference engines have combined, allowing the same computing power cluster to handle requests that are dozens of times what they could manage two years ago. When marginal costs are low enough, free closely resembles the logic of Google Search: free entry, monetization through the ecosystem.

The costs are equally clear. In the first quarter of 2026, OpenAI's revenue was 5.7 billion dollars, with a non-GAAP operating loss rate of -122%, losing 1.22 dollars for every 1 dollar earned, predicting a net loss of 14 billion dollars for the year. Among the 1 billion weekly active users, there were 50 million paid subscribers, with a payment rate of about 5%.

Subscription fees clearly won't support this company; the money must come from elsewhere: advertising revenue, which reached an annualized income of 100 million dollars just 6 weeks after launch; continuous expansion of enterprise API, with Codex serving 5 million users per week; enterprise clients currently contribute over 40% of revenue...

In other words, ChatGPT's business model is shifting from "selling model subscriptions" to "collecting platform taxes": the model itself is free, while advertising, enterprise services, and the developer ecosystem built on top of the model are revenue sources.

The 1 billion weekly active users are the core of this strategy; it doesn't require every user to pay, just the need for users to open it daily, and then charge for a small number of high-value requests. This is a very classical internet platform economics; previously often compared with Claude's "enterprise market, pay for programming" route, it now shows that OpenAI is determined to create a universal entry point for the AI era.

DeepSeek: The servers can't keep up

DeepSeek's situation is completely different from OpenAI's.

V4 Flash topped the global call volume weekly chart on OpenRouter, processing 72.2 trillion tokens in a single week. According to OpenCode data, on August 1 alone, V4 Flash's daily processing volume reached 80 trillion tokens. During peak working hours, there were frequent timeout errors and slowdowns, leading to the introduction of peak and valley pricing mechanisms in mid-July (doubling prices during peak times), which was followed by a direct announcement of a substantial price increase on August 6.

In other words: There are too many users, and insufficient computing power.

Excessively low pricing attracted a large number of low-frequency, low willingness to pay requests, with server resources being occupied by ineffective requests, while businesses and developers who really needed deep reasoning were unable to achieve a stable experience. DeepSeek needs to block users who treat AI as a toy at the door and retain those willing to pay for high-quality reasoning.

Doubao launched a paid version on June 24 (with tiers of 68/200/500 yuan per month, with basic functionalities remaining free) as an action that follows the same logic. With 345 million monthly active users and daily inference costs in the millions, e-commerce commissions cannot compensate.

Now, the domestic large model competition has reached a fever pitch, with a cooling financing environment, and the lack of computing power continues to be a constraint. Past strategies relying on capital for survival and unlimited subsidies are no longer viable; proving that the large model business can be self-sustaining has become a more urgent task than demonstrating technical leadership.

Of course, the price increases for domestic large models do not mean abandoning pricing advantages. More accurately, the cold start phase of the price war is over. The task in the previous phase was "to make users access it," and now the task has shifted to "making users willing to pay for quality."

One sinks down, one rises up

When GPT-4 was first released, everyone was comparing the model capabilities; who had read more books, who scored higher on tests. By the time of GPT-5.6 Luna, the gaps between top models are visibly narrowing, and the marginal returns of continuing to compete over benchmark test scores are increasingly diminishing. Competition is shifting from "who is smarter" to "who is indispensable."

OpenAI has chosen to push free services at this juncture, betting on user engagement and entry status. It wants to become the default way for most people to engage with AI, just as Google once was the default for search.

Chinese large model manufacturers choosing to raise prices or launch paid versions are seeking certainty in their business closed-loop under the dual pressures of capital retreat and computing power regulation, no longer satisfied with a situation of "a lot of users but not profitable," and they need to prove that the large model business can generate revenue independently.

The directions are opposite, yet the anxieties are symmetrical.

OpenAI needs to prove before the funding window closes that platform taxes can cover inference costs; Chinese companies need to turn user scale into sustainable income before the peak of computing power is reached.

This differentiation may ultimately form a three-layer structure.

The lowest layer consists of everyday conversations and general Q&A. The model capabilities have already overflowed, and costs are low enough to ignore; free will become the norm. Whoever charges here will be abandoned by users; OpenAI's choice to offer Luna for free aims to pull as many people as possible into this layer.

The middle layer focuses on professional reasoning. Scenarios like code generation, data analysis, legal assistance, and medical diagnosis have rigid requirements for accuracy and depth, and users are willing to pay for results, with DeepSeek and Doubao's paid versions targeting exactly this layer.

The top layer involves agent execution. AI is no longer just about answering questions; it directly completes tasks for users: negotiating business, fixing code, passing audits... At this layer, the charging model will shift from "charging by tokens" to "charging by outcomes."

Each has its troubles, each has its dreams.

OpenAI sinks down, betting on entry and habits, trying to make 1 billion people reliant on it; Chinese large models rise up, trying to get users to actively spend money... Both dreams are not cheap, and the only certainty is: The era of cost-free services is over; the remaining free offerings are all calculated business ventures.