How Does ChatGPT's New Browsing Update Access the Internet and Plugins? Examples With Bard and Bing

TL;DR

ChatGPT's new browsing update lets ChatGPT Plus users access current internet information and use more than 70 third-party plugins. The demonstrated GPT-4 browsing workflow searches pages, reads selected sources, and produces summaries with numbered hyperlink citations, although it can be slow and was limited to 25 messages every couple of hours. Read on for the setup, examples, comparisons, research benefits, and limitations.

Transcript

so chat TPT just released a massive update here you can see from openai they said we're rolling out web browsing and plugins to all chat DBT plus users over the next week moving from alpha to Beta they allow chat gbt to access the internet and use the 70 plus third-party plugin so this is a major major announcement in the world of AI because of cou... Read More

Key Insights

• Browsing targets newer information: The central improvement was internet access for material beyond ChatGPT's stated 2021 information boundary. The Google I/O 2023 prompt deliberately tested a recent event rather than relying on older knowledge. This made the example a direct test of whether browsing could locate, read, and summarize timely web material.
• Access depended on the model: Browsing appeared inside the GPT-4 selection menu, while GPT-3.5 was identified as the option available to free users. The demonstrated workflow therefore began by selecting GPT-4 and choosing browsing instead of the default mode. The announced rollout specifically concerned ChatGPT Plus users.
• Plugins expanded the update: Web access was not the only announced capability. OpenAI's update also allowed ChatGPT to use more than 70 third-party plugins. The presentation treated browsing and plugins together as a significant expansion of ChatGPT's practical capabilities as the features moved from alpha testing into beta availability.
• Dates sharpen recent-event prompts: The presenter explicitly advised including the date when requesting a recap of Google I/O 2023. A date gives the browsing system a clearer target when searching for a particular recent event. In this demonstration, that specificity helped distinguish the intended conference and its announcements from less relevant material.
• The process remained visible: During browsing, the interface displayed searches, attempted clicks, page reading, information extraction, and completion. The presenter contrasted this visible activity with Bard and Bing, which delivered their responses immediately without showing the same process. Seeing the steps made it easier to understand which pages ChatGPT tried to use.
• Failed clicks were recoverable: One browsing attempt displayed multiple failed clicks, creating the impression that the task might not succeed. ChatGPT continued searching, opened other results, read available material, and eventually finished. The example suggests that an unsuccessful link did not necessarily end the broader research process when alternative sources remained available.
• Repeated prompts produced differences: The exact same request generated different browsing behavior on separate attempts. One response searched more content and supplied stronger citations, while another relied on a narrower route through the Google Developers Blog. The presenter considered the earlier, more extensively sourced generation better despite using identical prompt text.
• Inline citations improved usability: The favored response attached sources as small numbered hyperlinks near the supported information. It did not merely place an undifferentiated list of links at the end. This structure made individual claims easier to connect with articles and gave users material they could inspect when preparing research, blogs, or arguments.
• Source breadth affected quality: The presenter valued the run that explored more pages, including The Verge and Google search results. Its broader search process led to what was judged a better response with useful references. The comparison within ChatGPT itself showed that answer quality could vary according to which sources the browsing session successfully reached.
• Research was a strong use: The demonstrated citation format was positioned as particularly effective when assembling a research paper, writing a blog, or supporting an argument. ChatGPT could combine a summary with links to underlying articles. That pairing offered more practical value than an unsupported recap because readers could follow the cited material.
• Speed remained the tradeoff: GPT-4 browsing was described as remarkably slow, especially when compared with tools that returned answers immediately. The delay came alongside visible searching, clicking, reading, and extraction. For users trying to assemble a research paper within a few hours, this slower pace could become a meaningful constraint.
• Usage limits restricted throughput: The stated GPT-4 allowance was 25 messages every couple of hours. That ceiling mattered because browsing research may require repeated prompts, refinements, and follow-up questions. Even when the resulting summaries and citations were useful, the limit reduced how much work a user could complete rapidly in one research session.

Questions & Answers

Q: What are the key features of ChatGPT's new browsing update?

The update gave ChatGPT Plus users web browsing and access to more than 70 third-party plugins. OpenAI was moving these capabilities from alpha to beta and rolling them out over the following week. Browsing let GPT-4 search for current information, open pages, read their contents, and synthesize a response with source links. This mattered because the presentation identified ChatGPT's inability to access information newer than 2021 as a major previous limitation.

Q: How do you access browsing in ChatGPT?

The demonstrated interface offered GPT-3.5 and GPT-4 as separate model choices. GPT-3.5 was described as available to free users, while clicking GPT-4 opened a dropdown menu. That menu contained the ordinary default mode and a separate browsing option. Selecting browsing activated the web-enabled model, whose search and page-reading activity then appeared after a prompt was submitted.

Q: How did ChatGPT browse for information about Google I/O 2023?

The presenter asked it to recap Google I/O 2023 and included the date in the request. ChatGPT searched the web, attempted to open results, read accessible pages, and extracted material for its answer. In one run it used the Google Developers Blog, while another explored The Verge and other Google search results. It then combined what it found into a summary, demonstrating how browsing handled an event that occurred after the stated 2021 information boundary.

Q: Why should a browsing prompt include a date?

The presenter specifically recommended including the date when requesting the Google I/O 2023 recap. The date clarified which recent event the request concerned and gave the browsing system a more precise search target. That specificity was important because the purpose of the test was to retrieve timely information rather than depend on older stored knowledge. In the example, ChatGPT used the dated request to search current pages and assemble an event summary.

Q: Does ChatGPT show what it is doing while browsing?

The demonstrated browser displayed its activity as it worked through the request. The viewer could see it searching, clicking links, reading information, extracting relevant content, and finishing the browsing stage. Some attempts were marked as failed clicks, while successful ones led to pages such as the Google Developers Blog or The Verge. The presenter contrasted this visible workflow with Bard and Bing, which were described as providing responses immediately.

Q: How does ChatGPT browsing handle failed links?

A failed click did not necessarily stop the task. In one run, several clicks failed before ChatGPT successfully opened other pages and read their contents. It continued through Google search results, gathered available information, and completed the response. This showed why allowing the browser time to work could matter, even when its early activity looked unsuccessful or underwhelming.

Q: How does ChatGPT browsing compare with Bard and Bing?

ChatGPT browsing was presented as detailed and capable of showing the search process, but also remarkably slow. Bard was characterized as faster and efficient, while Bing produced quick results. ChatGPT's stronger example explored several sources and embedded numbered hyperlink citations near relevant claims. The comparison therefore framed the choice around priorities: immediate output favored the quicker tools, while visible exploration and detailed citation-supported responses were strengths of ChatGPT.

Q: What are the limits of GPT-4 browsing for research?

The two emphasized constraints were response speed and message availability. GPT-4 was said to allow only 25 messages every couple of hours, and browsing could take considerable time while it searched and read pages. Those restrictions could impede someone trying to assemble a research paper within only a few hours. Even so, the resulting inline citations and source-backed summaries remained useful for papers, blogs, and arguments when the user could tolerate the slower workflow.

Summary & Key Takeaways

• Introducing the browsing update: OpenAI announced that web browsing and plugins were rolling out to all ChatGPT Plus users over the following week, moving the features from alpha to beta. The update allowed ChatGPT to access the internet and use more than 70 third-party plugins. This addressed a major limitation identified in the presentation, namely that ordinary ChatGPT could not access information more recent than 2021. The update was therefore presented as a major development for current-information tasks.

• Finding browsing in GPT-4: The interface showed two model choices, GPT-3.5 for free users and GPT-4. Selecting GPT-4 opened a menu containing the default mode and a separate browsing option. The presenter warned that browsing could initially feel underwhelming, but said it worked effectively when given enough time. The demonstration was designed to make its behavior visible and to support a comparison with Google Bard and Microsoft Bing, which already used GPT-4.

• Testing a recent event: The presenter asked ChatGPT to recap Google I/O 2023, a recent event where Google showcased generative AI tools, phones, software, and other launches. Including the date was recommended so the browser could target the correct event. After submission, the interface identified web browsing as the active model. ChatGPT then searched, opened links, read material such as the Google Developers Blog, extracted information, and returned what the presenter considered a decent summary.

• Observing variable search paths: Repeating the same prompt did not always produce the same browsing process or answer. One run searched more pieces of content, encountered several failed clicks, read material from The Verge and other Google search results, and ultimately completed the task. Another run used fewer sources. The presenter initially doubted the failed-click attempt would work, but it recovered and produced a response. This variability showed that identical prompts could lead ChatGPT through different sources and browsing routes.

• Comparing value and limitations: The strongest ChatGPT result included numbered hyperlink citations placed alongside relevant claims rather than collecting every link at the bottom. That format was useful for research papers, blog writing, and sourced arguments. However, GPT-4 browsing was described as remarkably slow and subject to a limit of 25 messages every couple of hours. Existing comparisons characterized Bard and Bing as quicker, while ChatGPT offered detailed responses and visible source exploration, making the best choice dependent on the task.