Model Slug
A model slug is the exact technical short designation used to address a specific version of an AI model, for example "gpt-4o-mini" or "claude-sonnet-4-20250514". It is not a marketing name but a unique address: a single wrong character leads to an error message or to the wrong model.
Large providers give their AI programs two names. One is the marketing name that you read about in the press. The other is a terse string of lowercase letters with hyphens, often ending in a number or a date. This second designation is called a model slug. A program that automatically talks to an AI service must specify exactly this slug, otherwise the service does not know which of the many available programs is meant. Examples are “gpt-4o-mini”, “gemini-2.5-pro” or “claude-sonnet-4-20250514”.
Why a single letter decides the bill
Providers rarely offer just one model; often they offer ten or more at the same time. They differ greatly in price, speed and capability. A small model can be twenty times cheaper per request than the largest one. Anyone who accidentally stores the slug of the expensive model in their own program only notices it on the monthly invoice. For companies, the slug is therefore a hard-nosed cost decision, not a formality.
Then there is traceability. If a research team claims that a model passed a test, the statement is almost worthless without the slug. The marketing name stays the same for months while several versions are swapped out behind it. Only the slug with a date reveals which version was actually tested. Serious comparison tables and technical articles always include it for exactly this reason.
A third point concerns stability. Providers eventually shut down old models; this is called deprecation. Such announcements are made via the slug, not via the marketing name. Anyone who ignores these announcements will one day find their application receiving nothing but error messages.
How such a string is structured
A slug usually follows a fixed pattern made up of several parts separated by hyphens. At the front is the model family, such as “gpt” or “claude”. Then comes the generation as a number, indicating which major revision it is. Often a size designation follows, such as “mini”, “flash” or “opus”. At the end there is frequently a date in year-month-day format, marking the exact release state.
You can picture this like the designation of a car. “Golf” is the family, “8” the generation, “GTI” the trim level, the model year the exact version. Nobody just orders “a Golf” from the dealer if the price is supposed to be right. Just as little does a program get by with simply specifying “Claude”.
Capitalization matters: slugs almost always consist only of lowercase letters, digits, periods and hyphens. Spaces are not allowed. There are also so-called alias slugs without a date, such as “claude-sonnet-4”. They automatically point to the respective latest version. This is convenient but risky: the behavior of your own application can then change overnight without anyone having changed a setting.
Where you run into slugs
Slugs most commonly appear in the providers' technical documentation. There you find tables listing all available models, the respective price per amount of text, and the shutdown date. Anyone writing a program that addresses an AI via an interface copies the slug directly from there into their own code.
They also turn up in rankings and test reports. When a news site reports that a model beat the competition, the slug is usually found in the footnote or in the table column. Discussions in developer forums run almost exclusively on slugs, because marketing names would be too imprecise there.
A common misconception is that the slug is simply the marketing name in lowercase letters. This is often not true. Some products carry a completely different name to the outside world than the model that actually works internally. Anyone who wants to know what a service really uses must look up the slug, not read the marketing material.
