OpenAI Moves Downmarket, DeepSeek Moves Upmarket

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

The day before, numerous DeepSeek API users received a notification: The company plans to raise API service prices across the board in the near future, with “expected significant increases.” Specific prices and implementation timelines will be announced later.

This seems somewhat counterintuitive.

Over the past two years, the narrative most favored by Chinese large model vendors has been extreme cost-effectiveness. Previously, DeepSeek provided performance close to GPT-4 at less than one-thirtieth the price of OpenAI, earning it the industry moniker “Price Butcher,” while Liang Wenfeng became known as “Saint Liang.” The consensus was: The core advantage of domestic large models is driving the cost of using AI down to rock bottom.

Now, that floor is giving way. OpenAI has become the entity giving things away for free to the world, while Chinese vendors are starting to discuss payments and price hikes. Has the balance of offense and defense truly shifted?

OpenAI: Free Is Not Charity

Let’s clarify one thing first: GPT-5.6 Luna is not OpenAI’s most powerful model. It is positioned in the mid-tier capability range, more than sufficient for handling daily conversations, simple writing, and basic translations, but complex reasoning and multi-step code analysis still rely on the higher-end SOL series. OpenAI is using a “good enough” model to cover the broadest user scenarios.

This strategy comes with both confidence and costs.

The confidence stems from the cost side. Over the past 18 months, the unit token cost curve for large model inference has dropped sharply: optimizations in model architecture, maturity of quantization techniques, and upgrades to inference engines have combined such that the same computing cluster can now handle dozens of times more requests than it could two years ago. When marginal costs are low enough, going free approaches the logic of Google Search: free entry, monetizing the ecosystem.

The costs are equally clear. In Q1 2026, OpenAI’s revenue was $5.7 billion, with a non-GAAP operating loss margin of -122%. For every $1 earned, the company lost $1.22, with a projected annual net loss of $14 billion. Of the 1 billion weekly active users, 50 million are paid subscribers, representing a conversion rate of approximately 5%.

Subscription fees alone clearly cannot sustain the company; the money comes from elsewhere: the advertising business generated $100 million in annualized revenue within just six weeks of launch; the 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 switching from “selling model subscriptions” to “collecting platform taxes”: the model itself is free, while the revenue sources are the advertising, enterprise services, and developer ecosystem built on top of it.

The 1 billion weekly active users are the core of this strategy. It does not need every user to pay; it only needs users to open it daily, charging only for a minority of high-value demands. This is a very classical internet platform economics model, previously often compared to Claude’s “enterprise market, paid coding” route. Now, it appears OpenAI is determined to become the universal entry point for the AI era.

DeepSeek: Servers Can’t Keep Up

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

V4 Flash topped the global weekly call volume charts on OpenRouter, processing 7.22 trillion tokens in a single week. According to OpenCode data, V4 Flash processed 8 trillion tokens in a single day on August 1 alone. During peak hours on weekdays, interface timeouts and lag occurred frequently. A peak-valley pricing mechanism (doubling prices during peak hours) was introduced in mid-July, and on August 6, a comprehensive and significant price increase was directly announced.

Put another way: Too many users, insufficient compute power.

Excessively low pricing attracted a large volume of low-frequency, low-willingness-to-pay calls, squeezing out server resources with invalid requests, while enterprises and developers truly needing deep reasoning failed to receive a stable experience. DeepSeek needs to keep users who treat AI as a toy out, while retaining those willing to pay for high-quality reasoning.

Doubao launched its paid version on June 24 (with three monthly tiers of 68/200/500 RMB, keeping basic features free), following the same logical line. With 345 million monthly active users consuming tens of millions of RMB in daily inference costs, e-commerce commissions cannot cover the deficit.

Today, the domestic large model competition has intensified, and the financing environment is cooling. Insufficient compute power remains a tight constraint. The previous strategy of surviving on capital and infinite subsidies is no longer sustainable. Proving profitability has become a more urgent task than proving technological leadership.

Of course, price hikes by domestic large models do not mean abandoning price advantages. A more accurate description is that the cold-start phase of the price war has ended. The task in the previous phase was “to let users access the technology”; the current task has become “to make users willing to pay for quality.”

One Moves Downmarket, One Moves Upmarket

When GPT-4 was first released, the comparison focused on model capabilities: who had read more books, who scored higher on exams. At the GPT-5.6 Luna stage, the gap between top-tier models is visibly narrowing, and the marginal benefit of continuing to compete on benchmark scores is diminishing. Competition is shifting from “who is smarter” to “who is more indispensable.”

OpenAI chose to launch free access at this juncture, betting on user engagement time and entry-point status. It aims to make itself the default way for most people to interact with AI, just as Google was once the default for search.

Chinese large model vendors choosing to raise prices or launch paid versions are seeking certainty in their business loops under the dual pressure of receding capital and compute controls. No longer satisfied with the situation of “many users but no profit,” they need to prove that the large model business can generate its own cash flow.

The directions are opposite, but the anxiety is symmetrical.

OpenAI must prove that platform taxes can cover inference costs before the financing window closes; Chinese companies must transform user scale into sustainable revenue before compute capacity peaks.

This divergence may ultimately evolve into a three-layer structure.

The bottom layer consists of daily conversations and general Q&A. Model capabilities have already exceeded demand, and costs are low enough to be negligible. Free access will become the norm. Whoever charges here will be abandoned by users; OpenAI making Luna free is an attempt to pull as many people as possible into this layer.

The middle layer is professional reasoning. Code generation, data analysis, legal assistance, and medical diagnosis are scenarios with rigid requirements for accuracy and depth, where users are willing to pay for results. DeepSeek and Doubao’s paid versions primarily target this layer.

The top layer is Agent execution. AI no longer just answers questions but directly performs operations for users: negotiating business, fixing code, passing audits... At this level, the charging model will shift from “pay per token” to “pay per outcome.”

Each has its worries, each has its dreams.

OpenAI moves downmarket, betting on entry points and habits, wanting 1 billion people to become dependent on it; Chinese large models move upmarket, wanting users to actively pay... Neither dream is cheap. The only certain thing is: The era of free access regardless of cost is over; the free access that remains is a calculated business.

Risk Warning and Disclaimer

The market carries risks; investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this content is at the user’s own risk.