OpenAI moves downward, DeepSeek climbs upward

Original Author: Xiaobing

On August 7, OpenAI announced that ChatGPT had reached 1 billion weekly active users, while also making GPT-5.6 Luna available to free users with unlimited text conversations.

The day before, multiple DeepSeek API users received notifications: The company plans to significantly raise API service prices in the near future, with "an expected substantial increase." Specific pricing and implementation timing will be announced later.

This seems somewhat counterintuitive.

Over the past two years, the most compelling narrative from Chinese large model companies has been extreme cost-effectiveness. DeepSeek previously offered performance close to GPT-4 at less than one-thirtieth of OpenAI's price, earning the title "price killer" in the industry, with Liang Wenfeng becoming "Saint Liang." The consensus was: The core advantage of domestic large models is driving the cost of AI usage to rock bottom.

Now that floor is shifting. OpenAI has become the one giving things away for free to the world, while Chinese companies are starting to talk about paid tiers and price increases. Between offense and defense, have the positions truly changed?

OpenAI: Free Isn't Charity

Let's get one thing straight: GPT-5.6 Luna is not OpenAI's most powerful model. It sits in the mid-tier capability range, more than sufficient for daily conversations, simple writing, and basic translation, but complex reasoning and multi-step code analysis still rely on the more advanced SOL series. OpenAI is using a "good enough" model to cover the broadest user scenarios.

This strategy has both confidence and costs.

The confidence comes from the cost side. Over the past 18 months, the unit Token cost curve for large model inference has been strikingly steep: model architecture optimization, maturation of quantization techniques, and inference engine upgrades have compounded, meaning the same compute cluster can now serve dozens of times more requests than two years ago. When marginal costs are low enough, "free" resembles the Google Search logic: free entry, monetize the ecosystem.

The costs are equally clear. In Q1 2026, OpenAI generated $5.7 billion in revenue with a non-GAAP operating loss margin of -122%, losing $1.22 for every $1 earned, with a projected net loss of $14 billion for the year. Among the 1 billion weekly actives, there are 50 million paid subscribers, a conversion rate of about 5%.

Subscription fees clearly can't sustain this company; the money comes from elsewhere: the advertising business hit a $100 million annualized revenue run rate within 6 weeks of launch; enterprise API continues to expand, with Codex reaching 5 million weekly users, and enterprise customers currently contributing 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, but the advertising, enterprise services, and developer ecosystem built on top of it are the revenue sources.

The 1 billion weekly actives are the core of this playbook. It doesn't require every user to pay; it just needs users to open it daily, then monetize a minority of high-value needs. This is a classic internet platform economics model, often compared to Claude's "enterprise market, paid coding" route. As it stands, OpenAI is determined to become the universal gateway of the AI era.

DeepSeek: The Servers Can't Keep Up

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

V4 Flash topped the weekly global call volume charts on OpenRouter, processing 7.22 trillion Tokens in a single week. According to OpenCode data, on August 1 alone, V4 Flash handled 8 trillion Tokens in a single day. During peak workday hours, API timeouts and lag were frequent. In mid-July, a peak-valley pricing mechanism was introduced (doubling prices during peak times), and on August 6, a comprehensive and significant price increase was directly announced.

In other words: Too many users, not enough compute.

The excessively low pricing attracted a large volume of low-frequency, low-willingness-to-pay calls, with server resources being consumed by junk requests, while enterprises and developers needing deep reasoning got an unstable experience instead. DeepSeek needs to keep users who treat AI as a toy at bay, retaining those willing to pay for high-quality reasoning.

Doubao's launch of a paid tier on June 24 (with three monthly tiers at 68/200/500 yuan, keeping basic features free) follows the same logic. With 345 million monthly active users and daily inference costs in the tens of millions of yuan, e-commerce commissions can't cover the gap.

Now, the domestic large model competition has become fiercely intense, and the financing environment is cooling down. Compute shortages remain a constant constraint, and the past approach of relying on capital injections and unlimited subsidies is no longer sustainable. Proving the ability to make money has become a more urgent task than proving technological leadership.

Of course, domestic large model price increases don't mean abandoning the price advantage. More accurately, the cold-start phase of the price war is over. The previous task was "getting users to try it," now it's "getting users to pay for what's good."

One Goes Down, One Goes Up

When GPT-4 was first released, the competition was about model capability—who read more books, who scored higher on exams. By the GPT-5.6 Luna stage, the gap between top models is visibly narrowing, and the marginal returns from competing on benchmark scores are diminishing. The competition is shifting from "who's smarter" to "who's more indispensable."

OpenAI chose to go free at this juncture, betting on user time and gateway status. It wants to become the default way most people interact with AI, just as Google became the default for search.

Chinese large model companies choosing price increases or paid tiers are, under the dual pressures of capital retreat and compute restrictions, seeking certainty in closing the business loop—no longer satisfied with a situation of "many users but no profits," needing to prove that large model businesses can generate revenue independently.

The directions are opposite, but the anxieties are symmetrical.

OpenAI needs to prove that platform taxes can cover inference costs before the financing window closes; Chinese companies need to convert user scale into sustainable revenue before compute capacity hits its ceiling.

This divergence may eventually crystallize into a three-tier structure.

The bottom tier is daily conversation and general Q&A. Model capabilities have already reached saturation, costs are negligible, and free will become the norm. Anyone charging here will be abandoned by users. OpenAI making Luna free is precisely about pulling as many people as possible into this tier.

The middle tier is professional reasoning. Code generation, data analysis, legal assistance, and medical diagnosis—these scenarios have rigid requirements for accuracy and depth, and users are willing to pay for results. This is exactly the tier DeepSeek and Doubao's paid versions are targeting.

The top tier is Agent execution. AI no longer just answers questions but directly completes operations for users: closing deals, fixing code, passing audits... At this level, pricing models shift from "per-Token billing" to "per-outcome billing."

Each has its troubles, each has its dreams.

OpenAI is going down, betting on gateway status and habits, wanting a billion people to find it indispensable; Chinese large models are going up, wanting users to actively pay... Neither dream is cheap, and the only certainty is this: The era of free without counting costs is over. Whatever free offerings remain are those that make business sense.