ChatGPT was quietly losing. Most people didn’t know yet.

November 30, 2022.

Chatgpt was launched. 1 million people in five days. 100 million people in two months. It was the fastest-growing consumer service in history.

And from that moment on, everyone was mistaken.

“AI = Chatgpt”.

The numbers told a different story.

In February 2026, this happened in just a few days.

@OpenAI signed a contract with the Pentagon. The next day, ChatGPT app uninstallation jumped 295% in one day. One-star reviews exploded 775%. Five-star reviews evaporated 50%.

During the same period, @AnthropicAI quietly said, “We are not.”

The results were clear.

Claude U.S. Downloads Surge 88% per day (Appfigures). Top U.S. App Store. And on Feb. 28, just one day, Claude reversed ChatGPT daily downloads for the first time.

The moment the graphs intersect, quietly, was a historical event.

The app download was a facade. The flow of money was real.

Ramp Economics Lab analyzed more than 50,000 U.S. corporate credit card data. In fact, it was data showing where companies spend their money.

As of January 2026, the proportion of companies introducing Anthropic paid companies reached an all-time high of 19.5%. On the other hand, OpenAI fell slightly to 35.9%. A year ago, Anthropic was one-fifth.

API spending was more extreme. According to Menlo Ventures’ separate analysis of corporate API spending in December 2025, Anthropic reversed OpenAI (27 percent) with 40 percent. It was a reversal in the market where developers plug into real codes and pay for them.

Why did this happen.

It’s not just that Claude got better.

First, the purpose of use has changed.

Early users of Chatgpt were “tried it once.” Curious and curious. In 2026, the money companies spent on AI was “to actually make it work.” Code-making, document analysis, and agent operation. Claude’s tool-call stability and long contextual processing power made a decisive difference in this use.

Second, the criteria for trust have changed.

Chatgpt started out as a consumer product. Quickly, funny, broadly. Claude was built from the beginning with the goal of “safe and reliable AI.” When companies spend money, what they want is trust, not fun. Denial of the Pentagon contract proved this positioning to the market in just one day.

Third, the era of agents has arrived.

The core of AI in 2026 was not “one conversation.” It was an agent that completed the task autonomously. The agent should not be wrong. It had to be consistent. The tools had to be handled correctly. Claude was overwhelmingly chosen under these conditions.

The Korean market was a little different.

When it came to “AI” in Korea, it was still Chatgpt. No news, YouTube, company training. But the people who actually run agents, code, and implement automation have already gone to Claude.

Personally, this gap is likely to widen further in the future. The public perception I have seen has always been later than reality.

So we have to ask one question.

Isn’t it because the AI I use now is “became famous in 2022”.

Tools should be chosen for purpose; having a conversation and having an agent work autonomously 24 hours a day are other requirements.

The data have already shown directions.

The only difference left is whether to look in that direction or not.

For reference, I use Claude for the Dolsoe agent, but codex5.3 is used for coding and design.

Appfigures / Sensor Tower US app download data (Feb 2026) Ramp Economics Lab enterprise AI spending data (Jan 2026, 50,000+ enterprise) Menlo Ventures enterprise API spending analysis (Dec 2025) Personal analysis, not investment advice.