No. ChatGPT runs on OpenAI's servers, and the app on your phone or computer is only the way in. With no connection you get an error message, not an answer. OpenAI's help center has no page that describes an offline mode (we searched it on 25 October 2026), and the pages it does have treat a missing connection as a fault to fix.
What does work without a connection is a different kind of software: an AI model that you download once and run on your own laptop. OpenAI has published two of those itself, and a program such as Ollama can run them. In a blackout the laptop still needs power, the model needs enough memory, and in our own test small offline models got safety numbers wrong. The sections below take those in turn.
What OpenAI's own pages say
The help article on the mobile app says you may see an error that asks you to check your device's date and time or internet connection, and its fix is to check your internet connection, then close and reopen ChatGPT. The error-messages article says that when you see "A network error occurred.", the error suggests that your device is unable to establish a reliable connection to OpenAI's servers. OpenAI's network guide for company IT departments lists separate addresses for the iPhone app, the Android app and the desktop app (ios.chat.openai.com, android.chat.openai.com and desktop.chat.openai.com) and asks that they are not blocked. Apps that reach out to servers like that are not built to run without them.
That covers the questions people type most often: the iPhone app, the desktop app, a paid plan. We found no page that gives any app or plan an offline mode. We can only report what the help center's search returned on the day, and OpenAI can change its products at any time.
Two pages can mislead. One is titled "Offline web search for ChatGPT workspaces". It describes a setting for eligible workspaces that lets ChatGPT use OpenAI's indexed and cached web content instead of live web search at the time of each request. ChatGPT itself is still online. Only the way it searches the web changes.
The other is the "What is ChatGPT?" FAQ, which says ChatGPT is not connected to the internet. The next sentence is about limited knowledge of events after 2021, so the page is describing what the model knows. It does not say the app works offline.
Why a blackout makes it worse
Your router and modem need power too. The FCC warns that phone service delivered over a broadband line will not work during an electric outage without a battery or other backup power, and says you may be able to connect through secure Wi-Fi if it is available nearby, so the one thing ChatGPT needs is among the first things a blackout takes away.
What runs offline instead
A language model is a very large file of numbers, and a program on your laptop loads it and answers questions with no help from the internet. You need a connection once, to download the program and a model.
LM Studio's documentation says the app does not in general require the internet, and that once a model is on your machine it will run locally and you should be good to go entirely offline. Downloading models is the exception: it needs a stable (and decently fast) internet connection, so it has to happen before you need it. Ollama is another program for the same job. It notes that models can be tens to hundreds of GB in size, so check your free disk space.
One trap is worth knowing. Ollama also offers cloud models, which it describes as models run in Ollama's cloud. A model with "cloud" in its name, such as the cloud versions listed next to gpt-oss, stops working the moment your connection does. Ollama's documentation says that to use only local models, you disable cloud features. Our guide to running AI offline on a Mac walks through the setup and the check with Wi-Fi switched off.
OpenAI's own open models
OpenAI released gpt-oss-120b and gpt-oss-20b on 5 August 2025 and says the weights are freely available to download on Hugging Face. It says the smaller one can run on edge devices with just 16 GB of memory, and the larger one on a single 80 GB graphics card. OpenAI's own cookbook has a guide for setting up gpt-oss-20b with Ollama, to chat with it offline. Ollama lists the 20b download at 14GB.
As of 25 October 2026, OpenAI's Hugging Face page shows no newer open model made as a general chat assistant. Its newer entries are two gpt-oss-safeguard models, which classify content, a privacy filter and a research model, and the rest are older speech and image models.
These models are not ChatGPT, and OpenAI itself describes open models as a complement to its hosted ones. Offline there is no web search either, so a model knows only what it learned in training.
Will it fit on your laptop?
Memory is the limit. LM Studio recommends 16GB or more on a Mac and says you may still manage on 8 GB with smaller models and modest context sizes. Our own test Mac had 8 GB, and the two models we ran took 2.0 GB and 3.3 GB on disk, nowhere near the 14GB of gpt-oss-20b. For a Mac with Apple silicon, our rule from the test is 16 GB of memory. On 8 GB a small model worked only with few other apps open, and a 4B model was already too slow.
What a small offline model got right and wrong
We ran two small models, Llama 3.2 3B and Gemma 3 4B, on an Apple M2 Mac with 8 GB, network off, through Ollama. Each read a small library of 24 short passages written from official advice, and we asked 33 survival questions about water, food in a power cut, gas, floods and first aid. The full write-up is in our offline AI test. It was one laptop and a small library, so read it as a sample.
Some of it went well. For all 27 questions the library could answer, the right passage was among the top 3 found, and first in 26. Llama began writing after a median of 2.4 seconds, and a full answer took about 4. Llama gave a fully correct answer to 26 of the 33 questions, Gemma to 22.
The mistakes were the kind that sound sure of themselves. Asked in Dutch how many drops of bleach go in 4 litres of water, Llama said 2 where the library says 8. Gemma said a closed fridge keeps food cold for 48 hours, where the USDA says a refrigerator keeps food safe for up to 4 hours. Llama added "don't call the fire brigade" to an answer about a gas smell, a line that appears nowhere in the library and is bad advice. Gemma suggested treating seawater with bleach, which does nothing about the salt. Counted up, Llama made 3 serious safety mistakes and Gemma made 4. Llama also declined only 5 of the 6 questions the library did not cover, and Gemma 3 of 6.
This is not a comparison with ChatGPT. We did not test ChatGPT or gpt-oss-20b, so we cannot say how a bigger model does. It may make fewer mistakes. It will still make some, and OpenAI's own help page says outputs may be inaccurate, untruthful and otherwise misleading at times. The lesson we take is about numbers. Copy a dose, a distance or a time from a source you can read, never from a model's sentence.
Getting ready, in order
- While you are online, ask ChatGPT for the checklists you want: water, food, heat, first aid. Check every number against an official page, then print them. Our water calculator and fridge and freezer timer are built on official figures for two of those.
- If you want an AI model on the laptop, check its memory first. Install it, switch Wi-Fi off and ask it something you already know the answer to.
- Plan the power. The laptop power station calculator turns battery size and laptop watts into hours.
- Keep paper copies, and a battery radio. The FCC suggests a battery-powered radio or portable television for emergency information during outages, which an offline laptop cannot give you.
We're building Waystone, an offline preparedness assistant that is still in development and not for sale. Because of the mistakes above, it is designed to show the page each answer came from, so you can read the source and not rely on the model's wording. If you would rather compare ready-made products than set up your own, our comparison of offline survival AI apps lists what each one says about itself, with prices checked on a date.