OpenAI 宣布终止对中国提供 API 服务,这会带来 (English)

Generated: 2026-06-24 12:43:28

---

Alright, sure. I read through this article twice, and the facts hold up—the data points and timeline line up. Now I'm going to take a pass at editing it, focusing on two things: fixing a few imprecise phrases, and scrubbing out that "content-farm" vibe so it sounds more like a real person talking.

---

The Day OpenAI Cut Us Off, I Actually Felt Relieved

It was the afternoon of June 25, and I was wrestling with a half-baked prompt when my email pinged.

I glanced at the sender and my gut dropped—OpenAI. Starting July 9, all API traffic from unsupported countries would be blocked.

I quickly clicked open the list of 188 countries and territories, scanning it three times. Mainland China? Not there. Hong Kong? Also not there.

To be honest, my very first thought wasn't "I'm doomed," but rather, "What was coming finally came."

You know, when you look at this together with what happened on June 13, it all makes sense. That day, OpenAI stuffed a heavyweight onto its board—former NSA director Paul Nakasone. The guy ran U.S. Cyber Command and headed the NSA; his whole thing is cyber warfare and digital defense. It's basically a neon sign: "I'm closing the door. Get ready."

Plus, the U.S. Treasury has been keeping the screws on technology investment in China, with AI and semiconductors right in the crosshairs. Even if OpenAI desperately wanted to keep Chinese developers, they can't twist the government's arm.

So why did I feel relieved?

Because I'm honestly sick and tired of needing to jump through hoops just to get work done.

Remember? Last year I paid over two hundred yuan and spent half a day to get some so-called "GPT-3.5 account." Within a month I was done. Proxies kept disconnecting, one day they wanted phone verification, the next day my payment failed… I threw a ton of time at it and the output wasn't even as good as what I could write myself. Back then the domestic models were just starting out—Wenxin Yiyan would go off the rails eight out of ten times when writing articles. I almost smashed my keyboard.

In February this year I decided to give domestic AI another shot. I tried five all at once: Tongyi Qianwen, Zhipu GLM-4, Doubao, Wenxin 4.0, and Kimi. Simple test: write an 800-word product description, find a bug in a Python code snippet, summarize a 3,000-word technical article, simulate customer service chat.

And the results were pretty interesting. On Chinese understanding and long-text processing, these domestic models were already on par with GPT-4. Zhipu GLM-4 writes code more reliably than GPT-3.5; Wenxin 4.0 is pretty creative with marketing copy, though logic sometimes slips; overall Tongyi Qianwen is the most well-rounded, and the free tier is generous.

I posted screenshots of my tests in a tech group, and the first thing a veteran said was, "Did they pay you to write a paid review?" I replied: "Try it yourself and see."

---

So with this cutoff, who's really getting grilled?

The ones hurting most, I think, are the "shell" operators.

A lot of so-called AI apps in China are basically a layer of paint over OpenAI's API. The core user experience and product logic all depend on GPT, and they just wrote a frontend. These companies made money easily before, but really they were just charging a "threshold fee"—you can't access ChatGPT? I'll get you in. You can't call the API? I'll do it for you. Clumsy, unstable, and prone to blowing up.

Now OpenAI slams the door, the core gets cut off, user data is in someone else's hands, and there's no Plan B. I know a founder of an AI writing tool who talked to me last week. He said he'd already migrated his backend to Alibaba's Bailian. The first three days were full of errors that nearly made him jump out of his skin, but now it's basically running. I asked about the cost—it's a fraction of what it was. Alibaba's migration solution is very close to OpenAI's API interface, so he was lucky.

Individual users? Almost zero impact. Anyone who was already using the GPT API for writing or experiments had already gone through the headache of getting an overseas IP, overseas phone number, foreign currency credit card. If they could get a key, they have a way to keep using it. As for people still using the ChatGPT web interface—ChatGPT has never been open in China; you were all using third-party channels or mirror sites anyway. Those people will probably just switch to domestic AI—they're used to freeloading, so they'll use whichever free one is available.

---

The moment OpenAI sent out that notice, domestic vendors jumped into action.

Zhipu AI directly rolled out a "moving plan": 150 million tokens (50 million for GLM-4 plus 100 million for GLM-4-Air), plus a 1:1 concurrent quota for high-usage clients.

Baidu was even more direct: flagship Wenxin model is free for the first time, with 50 million tokens, and if you're migrating from OpenAI, you get an equal amount.

Alibaba, meanwhile, hung a side-by-side benchmark of Tongyi Qianwen and GPT-4 right on their homepage, with a one-line conclusion: "Our ability is no worse, and the cost is a fraction of yours."

Talking about this now, I feel a bit emotional. Two years ago, would any domestic model dare to go head-to-head with GPT? When Wenxin Yiyan launched, I even wrote an article making fun of it for failing basic common sense questions. But the iteration speed of these models this year has been amazing, especially Zhipu GLM-4. Its basic capabilities are already on par with GPT-4, and its Chinese generation is even better.

But let's not get carried away: there's still a gap at the top end.

On GPT-4o's multimodality, reasoning depth, and code generation quality, domestic models still have a generation gap. Alibaba's Joe Tsai said that domestic AI will take at least two more years to catch up. I think that's a pretty realistic assessment.

But the question is: does the average person need 99% of GPT-4o's power to write a weekly report, make a PPT, or do some data analysis?

The domestic models are at 80% and they're free—that's already comfortable enough for everyday tasks.

---

Long term, this might actually be a good thing.

Bloomberg said OpenAI's cutoff would accelerate the shakeout in China's large-model industry. I agree with half of that.

Shakeout is inevitable—companies without core tech that rely on wrapping will die off. But it's not just the shakeout that's accelerating; it's also the use of domestic models in real-world scenarios.

Before, a lot of companies used OpenAI because "the big companies are using it, so if I use it I can't be wrong." Now they have no choice, and that's going to force them to focus on how to make domestic models work, instead of arguing about