Before decoding any label, we should understand why labels exist at all.

An AI model is a trained computer system that reads your input and produces an answer. Companies do not release just one model and stop. They release many, over many months, each a little different from the last. Some are bigger, some are faster, some are newer, and some are built for special jobs. Without a naming system, nobody could tell them apart.

Names and numbers solve several real problems at once. They help buyers pick the right model. They help developers write software that calls a specific model. They help researchers compare results fairly, because a result only means something if you know exactly which model produced it. And they help companies tell customers when an old model is going away.

Most model names are built from the same handful of parts, like the ingredients on a label.

Not every company uses every part, and each company has its own habits. But if you learn the common pieces, you can make sense of almost any name you meet. Here are the parts you will see most often.

flowchart LR
  A["Family: Claude"] --> B["Tier: Sonnet"]
  B --> C["Version: 5.5"]
  C --> D["Exact identifier: claude-sonnet-5-5"]
        

The first two parts of the name tell you the most about what to expect.

A family is the product line that groups related models together. All models called Claude belong to one family made by Anthropic. Other companies have their own families with their own names. The family name tells you the maker and the general style of the product.

Inside a family, models are often arranged in tiers, which are levels that trade off capability, speed and cost. In the current Claude lineup, the publicly described tiers include Haiku, Sonnet and Opus, with a higher Mythos tier above them. Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Fable 5.1 are the most recent publicly available models as described at the time of writing. Claude Sonnet 5.5 is described as a smart, efficient model for everyday use.

The number is the part people most often misread, so let us take it slowly.

In many model names, the number shows the generation, meaning the broad design era the model belongs to. A move from 4 to 5 usually signals a significant step, such as a new training approach or a larger leap in ability. A decimal such as 5.5 or 5.1 typically signals an improvement within a generation, a smaller step than a whole-number jump.

But here is the catch. Unlike the strict numbering that software engineers sometimes use, AI companies are free to choose their numbers. A decimal does not follow a universal rule. One company might use 0.5 steps for mid-generation upgrades, another might use 0.1 steps, and another might skip numbers entirely for marketing reasons. So the numbers are best read as the company’s own timeline, not as a measurement.

The same model can have two different names, one for people and one for computers.

When you chat with an assistant in an app, you see a friendly display name, such as “Claude Sonnet 5.5”. It has capital letters and spaces because it is meant for human eyes. When a developer builds software, they must type an exact model identifier, sometimes called a model string, such as claude-sonnet-5-5. This version uses lowercase letters and hyphens because computers need precise, predictable text with no spaces.

For the most recent publicly available Claude models, the identifiers follow a clear pattern: claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5 and claude-fable-5-1. Notice how the decimal point in the display name becomes a hyphen in the identifier. That is a formatting habit, not a change of meaning.

Here is a subtle idea that trips up even experienced developers.

A snapshot is a frozen copy of a model at a specific moment. It is usually marked with a date, and it never changes. If you call that exact snapshot today and again next year, you should get the same underlying model both times, which is vital for anyone who needs stable and repeatable behaviour.

An alias is a friendly shortcut name that points to a snapshot. Aliases are convenient because they do not require you to remember a date. The risk is that the company can later move the alias to point at a newer snapshot. Your software keeps using the same alias name, but the model behind it has quietly changed.

flowchart TD
  A["Alias name"] --> B["Points to a snapshot today"]
  A -.-> C["May point to a newer snapshot later"]
  D["Dated snapshot name"] --> E["Always the same frozen model"]
        

Extra words at the end of a name often matter more than the number.

A parameter is one of the adjustable numbers inside a model that stores what it has learned. More parameters generally means more capacity, but it does not automatically mean a better model. Training quality matters a great deal too. Large companies offering hosted models often do not publish exact parameter counts, which is why you will see size labels mostly on openly shared models.

Naming schemes are not carved in stone, and understanding why they shift will prepare you for future surprises.

Companies refine their naming as their product lines grow. A scheme that works for three models can become awkward for thirty. Sometimes a company changes the order of the parts, sometimes it adds a new tier, and sometimes it adjusts what the numbers mean. Early in an industry, names often change quickly because everything is still being figured out.

flowchart LR
  A["Few models: simple names"] --> B["More models: tiers added"]
  B --> C["Frequent updates: decimals and dates added"]
  C --> D["New top tier: ladder reshaped"]
        

Every company has its own habits, so skills from one do not transfer perfectly to another.

Because there is no industry-wide standard, each organisation invents its own scheme. Some name their models after generations and add letters for features. Some use poetic words for tiers. Some openly shared model families use a number for the generation, a figure for the parameter size, and a word such as Instruct for the purpose. The result is that the same number can mean quite different things depending on whose model it is.

Let us practise on a real set of names, using what is publicly described at the time of writing.

This exercise shows the method. Read the family first, then the tier, then the number, then any tag, and finally confirm the exact identifier in the official documentation. Notice that lineups change over time, so the specific names in this example may be replaced by newer ones in the future. The reading method will still apply.

Models do not live forever, and the version name is how you find out when yours is leaving.

When a company releases a newer model, it usually keeps the older ones available for a while, then announces that they will be retired. This process is often called deprecation, which means a model is officially marked as outdated and scheduled for removal. After the retirement date, requests to that model stop working. This is one of the biggest practical reasons to understand version names: if your software uses a specific version, you must know when to move on.

The same question can produce different answers from different versions, and that is why precision matters.

Different versions are trained differently, so they can vary in style, accuracy, length of answers and how they handle tricky requests. A prompt that works perfectly on one version might need adjusting on another. This is why careful teams record the exact model name next to every test result. Without that note, a result is like a measurement with no unit.

flowchart TD
  A["Same prompt"] --> B["Version A"]
  A --> C["Version B"]
  B --> D["Answer with one style and accuracy"]
  C --> E["Answer with a different style and accuracy"]
  D --> F["Record the exact version with every result"]
  E --> F
        

A few simple habits will protect you from the most common confusion.

These errors appear again and again. Spot them early and you will save yourself trouble.

They usually show the generation or release order within a company’s own scheme. Whole-number changes often suggest a bigger step and decimals a smaller refinement, but each company decides its own meaning.

Not always. It generally means newer, but the best model depends on your task, speed needs and budget. Test candidates on your own work.

The display name, like Claude Sonnet 5.5, is meant for people. The identifier, like claude-sonnet-5-5, is the exact text software uses to request that model.

A frozen copy of a model at a particular moment, often marked with a date. It does not change, which makes behaviour repeatable.

An alias is a convenient shortcut name that points to a model. The provider can later point it at a newer one, so the behaviour behind the same name may change.

They share the same underlying model, but Fable carries additional safety measures for biology, cybersecurity and LLM research and development. Different names signal a difference in how the model is protected and offered.

It usually signals an early or limited release. Claude Mythos Preview, for instance, is used by a small number of trusted organisations and is not open to the general public.

Tiers are not always updated at the same time, so each can carry its own version number. Always read the number that belongs to the specific model you are choosing.

It is commonly used with openly shared models and stands for billions of parameters, a rough indication of model size. Size hints at capacity but does not guarantee quality.

Always use the provider’s official documentation and announcements. Lineups, names and availability can change, so any written guide is only a snapshot in time.

AI model names are compact labels that pack several pieces of information together: a family that tells you who made it, a tier that hints at capability and cost, a number that shows where it sits on the company’s timeline, and tags that describe its status or purpose. Behind the friendly display name sits an exact identifier used by software, and sometimes a dated snapshot that stays fixed while friendly aliases can move. Because there is no universal standard, the same number can mean different things at different companies, and even within one family the tiers may carry different decimals. The safest approach is to read every part of the name, confirm details in official documentation, use fixed names when stability matters, and record the exact version beside every result.