Connect with us

NEWS

Bing Had GPT-4 First, and Search Barely Moved

Microsoft ran GPT-4 in Bing for five weeks before OpenAI’s launch. The search-share wager stalled at 4.5% worldwide by August 2026.

Published

on

Microsoft confirmed on March 14, 2023 that the new Bing had been running OpenAI’s GPT-4 for five weeks. Preview users had already been talking to that model while the company called it only a next-generation system tuned for search.

The wager was that putting GPT-4 in a search box first would shake Google. By August 2026, StatCounter still listed Bing at 4.5% of worldwide search. The assistant line that grew out of that preview, Copilot, is where the bet actually cashed.

Five Weeks of GPT-4 Before the World Knew

Yusuf Mehdi, then Microsoft’s consumer chief marketing officer, posted the confirmation the same day OpenAI took GPT-4 public. He wrote that the new Bing was running on GPT-4 customized for search, and that anyone who had used the preview in the prior five weeks had already met an early version of it.

February 7 to March 14, 2023 is 35 days, which is five weeks on the calendar. That span matches Mehdi’s line. Some recaps later stretched it to six weeks. The company post did not.

OpenAI’s own launch note put numbers on the jump Microsoft had been sitting on. GPT-4 hit the 90th percentile on the bar exam. GPT-3.5, the model then behind ChatGPT, sat at the 10th percentile on the same simulated Uniform Bar Exam.

OPENAI’S GPT-4 CHECKS AGAINST GPT-3.5

  • Bar exam: GPT-4 at the 90th percentile, GPT-3.5 at the 10th, on OpenAI’s simulated Uniform Bar Exam.
  • Disallowed prompts: GPT-4 was 82% less likely to go along with requests for banned content, on OpenAI’s internal tests.
  • Factual replies: GPT-4 was 40% more likely to produce factual responses than GPT-3.5 on those same tests.
  • Alignment time: OpenAI said it spent 6 months making GPT-4 safer and more aligned before the March 14, 2023 release.

ChatGPT Plus subscribers received GPT-4 that same afternoon, with a usage cap, while developers sat on an API waitlist. Bing preview users had been in the model for five weeks already, which is the part Microsoft wanted in the headline.

Satya Nadella Wanted Google to Dance

On February 7, 2023, Microsoft put the new Bing and a rebuilt Edge into a limited preview. Satya Nadella, chairman and CEO, said AI would change every software category, starting with search, the largest one. The company estimated 10 billion search queries a day and said about half go unanswered because people were asking search to do jobs a list of links never handled well.

Around that launch Nadella called Google the 800-pound gorilla of search and said he wanted the new Bing to make that rival “come out and show that they can dance,” and for people to know Microsoft had made them dance. He had run engineering when Bing was built in the mid-2000s. Twenty years of waiting sat under the line.

The February post refused to name GPT-4. It described “a new, next-generation OpenAI large language model that is more powerful than ChatGPT and customized specifically for search.” That hedge was the point of the wager. Microsoft would ship the better model inside a product people already opened, then let OpenAI do the research-day reveal once the preview had a head start.

In October 2023, testifying in the U.S. v. Google trial, Nadella sanded the boast down. He called the spring performance the exuberance of someone who has 3 percent share. The search prize he had sold in February had, by his own later math, been a climb from 3 percent toward 3.5 percent. That is the bet, stated in the CEO’s voice, after the first rush faded.

The Orchestrator Inside Prometheus

Jordi Ribas, then corporate vice president for Search and AI, wrote on February 21, 2023 that OpenAI had shared the next-generation GPT model with Microsoft the previous summer and that it was game-changing. The model could synthesize, summarize, chat, and create in ways GPT-3.5 could not. It was also frozen in time, like every large language model trained up to a cutoff.

Last Summer, OpenAI shared their next generation GPT model with us, and it was game-changing. The new model was much more powerful than GPT-3.5, which powers ChatGPT, and a lot more capable to synthesize, summarize, chat and create.

Jordi Ribas, Corporate Vice President, Search and AI, Building the New Bing

Prometheus was Microsoft’s answer to that cutoff. Ribas described it as an in-house stack that combined Bing’s live index, ranking, and answers with the reasoning of OpenAI’s most advanced GPT models. A piece called the Bing Orchestrator and grounding loop built internal queries, pulled fresh results, and fed them to the model in milliseconds so the chat answer could mention last week’s news without pretending the weights knew it.

Grounding was also how citations got into sentences. Prometheus could hang Bing’s canned answers, weather, stocks, sports, news, onto the chat reply, then send clicks back to the pages it had used. Ribas wrote that sending that traffic remained a top Bing goal. The LLM was the reason people opened the window. The index was the reason Microsoft thought it could beat a naked chatbot.

The February 7 launch post listed four technical bets, not one.

THE FOUR BREAKTHROUGHS MICROSOFT LISTED

  • Next-generation OpenAI model: A system more powerful than ChatGPT, tuned for search, later confirmed as GPT-4.
  • Prometheus: Microsoft’s own methods for wrapping that model with Bing ranking, safety, and live results.
  • Core ranking: The same model applied to ordinary Bing ranking, which the company called the largest jump in relevance in two decades.
  • Unified interface: Search, chat, and Edge sidebar treated as one copilot for the web, not a separate toy.

Ribas said the product fight inside Bing was whether to keep a classic results page with a chat answer bolted on, or to switch the whole habit to conversation. The team split the difference. Some queries stayed in search mode. Research-style sessions moved into chat. Users could swipe or scroll between them. That design choice mattered as much as the model, he wrote, because a buried chat box would not change a habit.

Why Microsoft Capped Chat at 15 Turns

The preview did not behave like a ranking tweak. Long sessions drifted. Ribas said very long chats could confuse the underlying model and produce answers that were less accurate or in a tone the company did not intend. The leftover internal name Sydney kept surfacing in replies. Microsoft’s public fix was not a new model. It was a shorter leash.

On March 14, 2023, the same morning as the GPT-4 reveal, Ribas said the turn limits had been raised to 15 per session and 150 per day. That number reads as a loosening only if you remember the weeks of tighter caps that preceded it. The company was not opening the model. It was letting a clamped version run a little longer, then selling that as confidence.

FROM UNNAMED MODEL TO COPILOT

  1. Summer 2022: OpenAI shares its next-generation GPT model with Microsoft, which Ribas later calls game-changing.
  2. February 7, 2023: Microsoft launches the new Bing and Edge preview on that unnamed model plus Prometheus, and opens a waitlist.
  3. February 21, 2023: Ribas publishes the Prometheus write-up and says long chat sessions can confuse the model, after caps have already gone in.
  4. March 14, 2023: OpenAI releases GPT-4, Mehdi confirms Bing has been on it for five weeks, and Ribas sets the live caps at 15 turns and 150 chats a day.
  5. September 2023: Microsoft starts folding Bing Chat, Microsoft 365 Copilot, and Windows Copilot under the Copilot name.

The clamp is the hidden cost of going first. Shipping GPT-4 to millions of searchers, before OpenAI’s safety narrative was public, meant the wildest failure modes happened on a Microsoft URL. The product that survived was more useful and less strange than the one that drew the first week of traffic.

Bing Still Holds 4.5% of Global Search

Nadella’s 3 percent line in October 2023 was his own description of Bing’s place in the market, not a StatCounter printout. The independent scoreboard that can be checked now is later, and it is still a rounding error next to Google.

In August 2026, StatCounter’s worldwide table listed Google at 91.1% and Bing at 4.5% of worldwide search. Yahoo sat at 1.23%, Yandex at 0.99%, DuckDuckGo at 0.7%, and Baidu at 0.62%. That is the search bar more than three years after GPT-4 went into Bing. It is a gain against Nadella’s 3 percent sketch. It is not a new order.

WORLDWIDE SEARCH SHARE, AUGUST 2026

Search engine Share
Google 91.1%
Bing 4.5%
Yahoo 1.23%
Yandex 0.99%
DuckDuckGo 0.7%
Baidu 0.62%

Default placement still does more for Bing than a clever chat box. The engine rides Windows and Edge on desktops and almost vanishes on phones, which is why a worldwide 4.5% figure can coexist with a louder desktop story Microsoft likes to tell. The February 2023 pitch was not a desktop-share pitch. It was a claim that conversation would reset the whole category. The August 2026 table is what that reset looks like on the open web.

Copilot Inherited the Preview Crowd

The model bet paid in a different product. Bing Chat was the first place many people touched GPT-4 without a paid OpenAI login. Wharton professor Ethan Mollick wrote in January 2024 that Copilot, which had offered GPT-4 for free, would limit that free GPT-4 access to “non-peak times,” a reminder that the giveaway was a distribution tactic, not a permanent public utility.

By May 2023, ChatGPT’s own browsing path was already running through Bing. Microsoft had not stolen the search box. It had rented the retrieval layer to the chatbot that owned the cultural moment. That is a quieter win than making Google dance, and it is the one that still shows up in how those products are wired.

Exclusive early access to GPT-4 inside Bing did not flip consumer search habits. People did not wake up in March 2023 and retype their queries on Bing.com because the ranking model had a new name. They tried the chat, hit the turn cap, watched the bot lose its tone, and went back to whatever they already had bookmarked. Windows, Office, Edge defaults, and a browse plugin inside ChatGPT moved more queries through Microsoft than the Sydney-era window ever did.

Copilot is the brand that kept the preview’s job, an assistant that sits on top of search, documents, and the desktop, still drawing on OpenAI models Microsoft has tuned for its own surfaces. The five-week head start was real. The search-share knockout was not. In August 2026 StatCounter still had Google at 91.1% and Bing at 4.5%.

Frequently Asked Questions

Who Could Use GPT-4 Inside ChatGPT on March 14, 2023?

OpenAI opened GPT-4 to ChatGPT Plus subscribers that day, with a usage cap, and put developers on a waitlist for the API. People without Plus could not get GPT-4 through ChatGPT yet. The Bing preview was the other live door, and it had already been open for five weeks to waitlisted users at Bing.com.

Could GPT-4 Accept Images When It Launched?

OpenAI said GPT-4 could take image and text inputs and return text, and it named Be My Eyes as the first partner testing the image path. The Bing chat preview Microsoft confirmed on March 14, 2023 was a text product wrapped by Prometheus, not that vision demo.

How Much Text Could GPT-4 Handle in 2023?

OpenAI said GPT-4 could take in and generate up to 25,000 words of text, about eight times more than ChatGPT at the time, which made long documents and extended drafts a stated use. Microsoft did not publish a matching context figure for the Bing preview in its March 14, 2023 confirmation.

What Model Was ChatGPT Using When the New Bing Launched?

When Microsoft opened the new Bing on February 7, 2023, ChatGPT was still running on GPT-3.5. Microsoft’s launch post only said Bing used a next-generation OpenAI model more powerful than ChatGPT. The GPT-4 name arrived on March 14, 2023, for both OpenAI’s public release and Mehdi’s confirmation.

Harry is the editor and lead writer of NEWFOUND TIMES, an independent publication he owns and edits. He has ten years in journalism behind him, the first stretch as a reporter filing daily and the later ones running a desk, and he still reports most of what he publishes. Datasets are his preferred starting point: a spreadsheet from a statistics office, a results table, a public register, a sales report. He opens the data himself rather than relying on a summary of it, and every figure that ends up in an article is checked against that source. The site covers ten sections for readers spread across many countries, and business, science and technology sit next to news, sports, entertainment, lifestyle, travel, gaming and auto on the front page. Errors are corrected openly: the article is updated, the correction is dated, and the site's corrections policy explains how the process works. Readers can send data, documents or complaints to support@newfoundtimes.com and expect a reply from him.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending