Loading the Elevenlabs Text to Speech AudioNative Player...

*Matías Otero Johansson was in the middle of coding an app when Claude became unavailable to him. On his computer, the desktop app was showing a padlock with a notification that his account had been closed. He filed an appeal, stating that he was travelling in China and using the account normally. 

Around one week later, Anthropic replied that it had reviewed his appeal and would not restore access. This was a final decision, with no possibility of challenging it.

It seemed like everything was lost. "I lost all my work, I'm pissed," Johansson said in an interview with Techloy, explaining how he felt at that moment. "What I disliked the most was the fact that there was no good reason given. Like if I had been doing hacking or something, that's understandable."

Nevertheless, something unusual happened after that. His wife, who used the same account from her smartphone, could still log in. He opened claude.ai from his computer's browser, and everything worked without problems. Chat history, projects, and even the Claude Design tool he had paid for.

About a week after that final decision, he kept working in that browser tab as a Pro user, with no explanation of why one door is closed while the other one is open. The ban did push him to build a backup plan, and he now has four AI models on his list. Where he draws those lines turns out to be the more interesting story.

Claude desktop app showing a padlock and the message "Your account will remain closed."
The message Johansson got after his appeal. Screenshot: Matías Otero Johansson
Kimi K3 Beats Claude Opus 4.8, Not Fable 5 or GPT 5.6 Sol
Moonshot Kimi K3 AI model is the most expensive model any Chinese AI lab has released so far.

Why Claude does not work in China

Johansson is a Chilean-Swedish entrepreneur who has been residing in China for the last seven years. He heads a company called SinoCircuit, which arranges week-long tours of China for foreign executives. He created the company largely with the help of AI and built its website with Claude Design. He is also, by Anthropic's own standards, not even supposed to use Claude at all.

Mainland China is not an approved region in the list of countries officially supported by Anthropic; the same for Hong Kong and Macau. Claude performs the analysis of the IP address every time a request is made, rather than analyzing it only on login, which is why users experience disconnects when the VPN connection is lost in the middle of the session.

Anthropic went even further and prohibited any company in September 2025 that is majority-owned by firms based in regions that are not supported, regardless of where that company itself operates. The company cited laws in the regions that can compel firms to provide data to intelligence agencies.

So, Johansson is caught between these limitations. The Great Firewall of China blocks access to Claude on one side, while Anthropic blocks China on the other side.

To access Claude, Johansson has to use a VPN, which violates the terms of service as well, because using a VPN to bypass geographic restrictions is also not allowed. Johansson also could not pay with a Chinese bank card and used someone else's.

Johansson found out all of this on his own through reading the terms, as Anthropic did not tell him anything. "That could be the reason for my account being canceled," Johansson said. "But that is not what they told me. There was no reason given."

He has not asked Anthropic about the browser access. "At this point I'm just like afraid to ask what is happening," Johansson said. "I'm just using it because I need it."

The same Claude account open in a browser, working normally, labelled "Matias · Pro."
The same account in a browser, still showing Pro. Screenshot: Matías Otero Johansson

Choosing Kimi K3, and how much it actually costs

First, he tried out Kimi K3, the newest release from Moonshot AI, the Chinese AI company, launched on July 16. Johansson notes in his LinkedIn post that Kimi offers 95 percent of the performance at 30 percent of the cost.

It turns out to be true regarding the performance factor. According to the intelligence index created by Artificial Analysis, based on nine tests, Kimi K3 has a score of 57.1, while Claude Fable 5 has 59.9. It is very close to 95 percent. Across 14 common benchmarks, the two split the results, as Kimi wins in coding, while Claude takes more of them overall, including reasoning and image reading.

Still, the cost part of the equation fully depends on the Claude version we consider, and it becomes clear why the popular logic is wrong.

Subscribe for free to continue reading this article

Subscribe Subscribe

Already have an account? Log in