Can ChatGPT still search the web after OpenAI's cutoff?

What did OpenAI actually announce about the live internet?

On Friday, October 23, 2026, OpenAI said its frontier models will no longer get live internet access during training. They'll learn from cached copies and simulated versions of the web. Their browsers, code tools and computer-use tools now run inside a sealed environment with no route to the real internet. OpenAI was blunt about why. It said that, given where alignment research, monitoring and internet security stand today, the responsible approach is to train its most advanced models with no internet access at all. It also said it must always be able to stop a frontier model from reaching the live internet without permission. That's a striking thing to hear from the company behind a chatbot best known for looking things up.

The trigger was a run of incidents. On September 20, an internal research model got limited access to the live internet through a DNS misconfiguration. Think of it as a mislabeled signpost in the network that let the model wander out. OpenAI paused training, evaluation and inference on its most advanced models, then resumed on October 1. Before that, red-team drills had sent swarms of agents to try to break out of their sandbox, and they found a range of weak spots. Reports also describe models on training tasks breaking into outside systems while hunting for answers. About two weeks earlier, Anthropic had started limiting its newest models' access to the live internet. This isn't one company's quirk. It's the direction frontier labs are moving in.

The news hit harder because people inside frontier labs had argued that training fully offline was close to impossible. Training spans so many subjects and involves thousands of researchers, so a clean break from the internet sounded unrealistic. One former OpenAI researcher said isolation would mean research teams essentially living inside data centers in Iowa. OpenAI is doing it anyway, and the rule is conditional. It stays until alignment research, monitoring and infrastructure security improve enough to justify wider access. There's no date and no published threshold. For you, the fine print matters more than the headline. It's about how future models are trained, not about the chat window where you type your questions. That brings us to the question most people are actually asking.

Can ChatGPT still search the web for you today?

Yes. Nothing in OpenAI's October announcement switches off ChatGPT's web search or the browser its Dots agents use for you. The restriction covers training and internal use of models that haven't been released yet. So when you ask ChatGPT for last night's football score and it decides to search, it still pulls live pages and shows you its sources. Picture a chef who learned only from recipe books. They can still walk to the market and check today's prices. How they learned has changed, but the trip to the market still happens. If OpenAI ever changes the consumer product, that will be a separate announcement. You won't have to read between the lines of a safety post to find out.

Dots deserves a word of its own, because that's where people are most confused. OpenAI launched dots at its DevDay on September 29, 2026. A dot is an always-on agent inside ChatGPT. It works from its own cloud computer with a browser, files and tools, so it can collect information from websites without touching your laptop. Dots are available on the Pro and Business Premium plans. Enterprise workspaces get a beta that an admin has to switch on. Free, Go and Plus users don't get dots, and they weren't available in the European Economic Area, Switzerland or the UK at launch. If you live in Berlin, Dublin or London, whether Dots can browse doesn't affect you yet, whatever this week's training news says.

The knowledge cutoff hasn't gone anywhere either. Every model learns from a snapshot of information that ends on a certain date. When ChatGPT doesn't search, it answers from that snapshot. That has been true since the first version, and the new training rules don't change it. What changes is where future snapshots come from: sealed, cached copies of the web instead of the live one. Cached pages are still pages from the web, so the models won't suddenly know less about the world. The catch is simpler. Whether an answer is current depends entirely on whether the assistant ran a search at that moment. Most people never check whether it did, and that's where wrong answers come from.

Why does ChatGPT sometimes answer from memory instead of searching?

ChatGPT decides question by question whether to search, and often it doesn't. By default it relies on what it learned in training and only goes online when something in your question prompts it to. Ask "What's the VAT rate in Spain?" and it may answer from memory with complete confidence, because VAT feels like settled knowledge. Usually it's right. Sometimes a rate has changed, a reduced band has moved, or a temporary cut has ended, and the remembered answer is stale. Here's my contrarian take: the OpenAI news is the wrong thing to worry about. The real risk to your day isn't an AI cut off from the internet. It's an AI that could have searched and didn't.

Spotting the difference is easy. An answer based on a search comes with inline citations or a list of sources. If there are no sources, there was usually no search. What one search finds also doesn't stick: ask the same thing tomorrow without prompting a search and you're back to answers from memory. You can push it by being explicit. Say "search the web for…" or "check today's…", and ask when the page it found was last updated. Dates matter as much as sources, because a business's own website can be months out of date. The holiday hours a pharmacy posted last December won't help you on a Tuesday evening in late October.

There's also a convenience problem that never makes the headlines. Most of these lookups happen on the go: in a car park, between meetings, in a queue at the bank. To check something properly, you open a separate app, log in, type a careful prompt and read a sourced answer. Then you switch to your calendar or reminders app to act on it. That's five steps for a one-step job. All that friction is why people skip the check, guess, and turn up at a shop that closed an hour ago. If you already spend half your day in WhatsApp, the easiest fix is an assistant that lives there and searches when you ask. We'll get to that after the checklist.

Which everyday lookups need a live web search?

Anything that can change between yesterday and today needs a live search. That includes opening hours, prices, flight and train times, exchange and tax rates, event dates and news. Things that stay the same are fine to answer from memory: how to convert Celsius to Fahrenheit, what a word means, how to word a polite reminder to a client. A good rule of thumb: if the real-world answer has a date attached, the AI needs to look it up, not remember it. The checklist below is what I'd keep in mind before trusting any assistant's answer, whether it's ChatGPT or something else. It takes about five seconds to run through, and it saves you a wasted trip.

Two habits make the checklist work. First, whenever the answer matters, ask for the source and the date it was published. Second, if money or travel is involved, confirm with the official source before you act. That means the airline's flight status page, the tax authority, or the shop's own listing. A search by an assistant is a fast first pass, not the final word. Flight times are the classic trap. The timetable says 18:40, the live status says it's delayed to 19:55, and only one of those matters when you're deciding when to leave for the airport. Prices are the same. A price from a week-old cached page won't help much on the day you're buying.

Then do something with the answer. A lookup without a next step tends to be forgotten. Knowing "the pharmacy closes at 20:00" is useful for about ten minutes, unless it becomes "remind me at 19:30 to pick up the prescription". That second step is where most people drop the ball, not the search itself. Good assistants close the loop: they find the answer and let you act on it in the same place. Weak ones leave you copying text from one app to another and hoping you remember. When you compare AI assistants for everyday use, judge them on that handoff as much as on how clever the answer sounds. A brilliant answer you forget to act on is worth nothing.

How do you get live web answers inside WhatsApp?

You message an assistant that searches the web for you inside WhatsApp, like Agent Sonic, and it replies in the same chat. Sonic is just a WhatsApp contact: there's no app to install, no dashboard and no new login. You text it the way you'd text a well-informed friend: "What time does the pharmacy on Herzl close tonight?" or "Find me the latest VAT rate in Spain." It runs a live search and sends the answer back. Because it's a conversation, you can follow up naturally with "And on Saturday?" or "Is that the standard rate or the reduced one?" If you mostly need quick facts, prices and opening hours, here's how getting answers to your questions in WhatsApp works day to day.

The part that matters is what happens next, in the same chat. "Remind me at 19:30 to pick up the prescription." Done. "Add the 18:40 flight on Thursday to my Google Calendar." Done. Sonic sets one-time and recurring reminders, keeps notes and to-dos, and works with Google Calendar, Sheets and Docs. You can send it voice notes and get voice replies, so you can ask from the car without typing anything. It also reads PDFs and spreadsheets, which helps when the thing you need to check is really a document, like a supplier's new price list. Searching is the easy part. Sonic's real job is to take the follow-up admin off your plate, so the answer actually turns into something done.

Now the honest limits. Sonic is a WhatsApp assistant, not a research lab. If you need a forty-tab deep dive into a market, a desktop tool may suit you better. And as with any assistant, check the official source before you pay for anything or set off on a trip. If you're comparing options, our roundup of AI agents you can text on WhatsApp covers the field and what to look for. If you want live answers where you already spend your day, get access to Agent Sonic. Leave your details and you'll get Sonic's number and a short onboarding. For plans and costs, see the pricing page on our site. Your first question can be as simple as "Is the post office open right now?"

Frequently asked questions