Why do the answers change when the name stays the same?
A service with the same name gives different answers because the model running inside it is replaced by a newer version. Names such as ChatGPT (OpenAI), Claude (Anthropic) and Gemini (Google) are service names, and the models inside have names of their own. When a developer releases a new model, the app usually switches to it without the user doing anything.
A new version answers the same question in a different way. The length and tone of its answers change, and it may miss a problem the old one solved. If you have built an LLM (large language model, a program that reads text and writes answers) into a work process, manage it by the model version, not by the service name.
Even the same version gives slightly different answers each time. That comes from a setting that adds variety to the answers, which is a separate cause from a change of version.
Read service names and model names separately
A service is a set of screens and features, and the model is the part inside it that writes the text. One service often lets you pick from several models. Ratings and prices belong to each model, so compare by model name.
Replacements do happen: on April 30, 2025, OpenAI replaced GPT-4 in ChatGPT with GPT-4o. A rating you read earlier may not be a rating of the model inside the service you use now.
Before reading ratings, check the name of the model you are using in the app settings or the model picker. Once you know the name, you can look up that model's ratings and release date on the use pages. If an app does not show the name in its picker, the developer's announcements or help pages often say when it switched to which model.
What do the numbers and dates in model names mean?
Model names often carry a number for the generation or a date for when the model was made. The number is a guide to the generation, and one generation can include models of different sizes and speeds. Do not judge old and new by the size of the number alone; look at the release date. A name with a date points to a version frozen on that day.
Developers sometimes offer both names that point to a fixed version and names that move to each new version. As of October 2026, Google sorts Gemini names into four types: stable versions, preview, latest and experimental. A stable version's name usually does not change, while a name with latest is swapped to the new version at each release. As of the same date, Anthropic explains that every Claude model ID points to a fixed version, and the model behind it does not change as long as you use that ID.
Read a name that ends in the characters preview or latest as one whose contents may change later. For work, choose a name that points to a fixed version.
The API lets you pin a version
In an app, the developer decides which model you use, but in the API (the gateway for using an LLM from a program) you specify the model yourself by name. If you specify a fixed version, the same model answers until the developer stops offering that version. The answers keep the same shape, which suits work such as classification and extraction that needs answers in the same form every time.
Fixed versions also come to an end. For each model, Anthropic lists a date until which it will keep offering the model at least. When the end draws near, compare old and new answers on your usual questions before moving to the next version.
When you pin a version, keep a record of the model name, the date you tested, and the answers at that time. If the answers go wrong later, you can tell whether the version changed or whether your instructions or the documents you passed changed. In an app too, saving an answer that caught your attention along with the model name and date lets you compare later.
In an app the model inside changes without notice, and in the API you change it yourself. Keeping this in mind makes it easier to find the cause when the answers change.
Why does the model not know about new events?
A model learns from text collected up to a certain date. That last date is called the training data cutoff. Events after the cutoff are not in the text it learned from, so the model does not know about them.
As of October 2026, Anthropic lists two dates for each model: the last date through which its knowledge is most extensive and reliable, and the last date of the range of data used for training. Read anything after the first date as a period the model may know only thinly. Each version has its own cutoff, so changing versions can also change what the model knows.
When an app searches the web to answer, it can also answer about things after the cutoff. When it does not search, it may make up a plausible answer about an event it does not know. When you ask about new rules, prices or news, open the sources attached to the answer and check them.
Summary - check the name and release date before comparing
When the answers change, first check whether the version of the model inside has changed. Read ratings by model name, not by service name, and pin a version in the API if you build the model into your work. When you ask about new events, consider whether they fall after the cutoff, and check the sources.
The tables on this site's use pages show the release date of each model. If several versions from the same developer appear on the page for building a chatbot, the release dates tell you which is newer. Whether a new version suits your work better than the old one is something to decide by putting the two side by side on your own questions.