From 2.5 billion queries to AI browsers: Can the "slow revolution" of ChatGPT Agent overthrow Google?

"Did you ask ChatGPT today?" When this question becomes as commonplace as "Have you eaten?", OpenAI quietly announced a set of figures that would make any product manager lose sleep: ChatGPT receives 2.5 billion user instructions per day, which amounts to 912.5 billion times annually, equivalent to 2,900 "dings" per second. If we imagine each question as a search, then ChatGPT already accounts for 18% of Google's annual search volume (5 trillion times), and this gap is narrowing at a visible pace.

What's even more intriguing is that on the same day OpenAI released the data, it introduced a new feature called ChatGPT Agent to Pro users - it can not only answer questions but also "take action". Thus, a covert battle about "from answers to actions" has begun.

Behind 2.5 Billion Questions: The "Water, Electricity, and Gas" Moment of AI

What does 2.5 billion questions per day mean? It has boosted ChatGPT's weekly active users from 300 million to 500 million within half a year, with a growth rate comparable to TikTok's most explosive year in 2018. US users contribute 13% of the questions but account for over 60% of OpenAI's paid revenue - the delicate balance between the paywall and free trial is laid bare by the data. Compared with Google's 14 billion daily searches, ChatGPT's questions are more "substantial": the average number of dialogue rounds is 3.7, and the average stay time is 4 minutes and 12 seconds, while the average session of a Google search is only 26 seconds. In other words, users are "chatting" rather than "searching".

When asking questions becomes a habit, the next step naturally is "let AI do it for me". This is the backdrop for the emergence of Agent.

ChatGPT Agent: An Intern on the First Day or a Future Super Assistant?

A reporter from The Verge spent $200 to subscribe to the Pro version and immediately put Agent into "actual combat":

1. Buy a Japanese retro lamp under $200. Agent spent 50 minutes searching, comparing prices, filtering, and adding items to the cart on Etsy, but in the end, it just put the products into a "virtual machine" shopping cart - when the reporter logged in, it was empty.

2. Order flowers for a friend in Colorado. Agent found four local flower shops and even provided the option of "delivery before Saturday", but at the last step, it said "Sorry, I can't place the order directly."

3. Ask Agent to log in to the bank account and set up automatic transfer. It directly received a red error message: "Sorry, I can't assist with such tasks."

Yash Kumar, the product leader at OpenAI, didn't shy away from this: "We are currently optimizing task completion rate, not latency." Isa Fulford, the research leader, added: "Even if it takes half an hour, it's still faster than doing it yourself." In a single sentence, Agent was positioned as a "back - end intern": you can go for a coffee while it slowly goes through the process, and when you come back, it will give you a report that may not be 100% usable.

But don't rush to laugh. Google Search in 2004 often returned 404 errors, and Siri in 2011 was mocked as "artificial stupidity". The first day of any infrastructure - level product is always in a mess. The key lies in the "possibilities" that Agent has opened up: for the first time, it combines Operator (web clicks) and Deep Research (multi - step reasoning) into the same model. It can view images, turn pages, write code, and run scripts in the terminal; it operates in a cloud - isolated container, and the local computer is just a "screen mirror". This means that in the future, even if you turn off your computer, Agent can help you finish annotating 100 pages of financial reports overnight; OpenAI has pre - installed a "monitoring mode" and "confirmation for irreversible operations" to keep high - risk tasks in finance, medicine, and law outside the guardrails - prioritize stability first, then speed.

The Browser War: AI to Become the New "Address Bar"

Reuters reported that OpenAI will release an AI browser based on Chromium in the "coming weeks". What will happen if we put Agent in the address bar?

When you enter "Help me book a two - person table at an Italian restaurant this Friday at 7 p.m.", instead of jumping to 10 pages, the browser will directly pop up a list of available reservations on OpenTable, marked with ratings, distances, and allergen menus.

When you open a financial report PDF, the Agent in the sidebar will drag the key data into Excel and generate a year - on - year comparison chart. When you see a post on Reddit saying "Please recommend a white sea - view computer case under 5,000 yuan", Agent will automatically compare prices on JD.com and Taobao and create a timeline of historical low prices, coupons, and today's flash sales.

To be more radical: if the default homepage of the browser is no longer Google.com but ChatGPT.com, Google's pay - per - click advertising model will be undermined. Chrome's moat is the "default search engine revenue sharing". Once Agent can bypass the search and directly reach the results, Google's money - making machine will lose a gear.

Three Suspenses of the Slow Revolution

1. The reliability gap: Currently, the success rate of Agent is about 75%. It may be acceptable for scenarios with high error tolerance like "buying flowers", but once it comes to high - risk fields such as finance, medicine, and law, a single mistake may lead to lawsuits. OpenAI needs to find a new Pareto optimum between safety and ability.

2. The economics of computing power: 2.5 billion questions per day have already cost OpenAI $4 billion in computing power annually. If Agent extends the average task duration to 30 minutes, the number of inference tokens will increase exponentially - unless the model efficiency is improved by an order of magnitude, the subscription fee will have to rise.

3. Power redistribution: Fidji Simo, the CEO of Instacart, is about to join OpenAI to lead the "applications" department. Her task is to find the "killer - app scenarios" for Agent, making AI not just a toy for geeks but a necessity for the general public, just like food delivery and ride - hailing services. In other words, whoever can define the interaction paradigm of the "super assistant" will hold the entrance ticket to the next decade.

Will the Next "Google It" Disappear?

In 2004, on the eve of Google's IPO, no one believed that a search box could subvert portals; in 2012, when WeChat launched official accounts, no one believed that chatting could replace browsers. Today, ChatGPT Agent is still clumsy, slow, and even a bit cute, but for the first time, it has shortened the path from "question to action" to the distance of a single sentence.

Maybe five years from now, we will miss the days of "comparing prices by opening ten tabs" just like we miss dial - up internet. And when children ask "What is a search engine?", we'll just shrug and say:

"Oh, that was the era when AI couldn't do things on its own."

This article is from the WeChat official account "Shanzi". Author: Rayking629. Republished by 36Kr with permission.