Home / What a language model is, without the jargon

One family of models, a choice between speed and depth

The models within the same family share the same base training and really only differ in the trade-off between how fast the answer comes and how deep the reasoning applied before answering goes.

Claude exists as four distinct models at the time of writing, Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5. All four come out of the same base training and are mainly distinguished by a single setting, the trade-off between how fast the answer comes and how deep the reasoning applied before answering goes.

A common foundation, a different slider

Choosing a model amounts to moving a slider, not changing tools. At one end, a model answers within seconds and suits short, repeated tasks, an acknowledgement of receipt, a rewording, a sort of messages. At the other end, a model takes longer because it chains more reasoning steps before concluding, which matters for a file that needs cross checking or a plan that needs verifying on several points. The documentation published by Anthropic recommends starting with Opus 5 for most uses, and reserving Fable 5.1 for tasks that demand demanding reasoning or that stretch across many chained steps.

A concrete example helps to feel the difference. Here are two requests the same volunteer might make on the same day.

Short request: rewrite this welcome message in three warmer sentences.
Long request: compare the twelve quotes received this quarter, flag the most
significant price differences and propose a reasoned ranking.

The first task is handled within seconds by a fast model. The second benefits from being given to a model that takes the time to cross check the figures before concluding.

The price follows the same trade-off

The price charged per million words exchanged varies strongly from one model to another, and this variation follows exactly the same axis as speed, the deepest model costs the most, the fastest costs the least. The figure below reproduces these four published prices, with their source and date, rather than copying them here where they would go out of date without anyone noticing.

Remembering the exact names matters less than remembering the principle: a deeper model is not a model that answers every request better, it is a model that answers better those requests that need several reasoning steps. The choice of model, covered here, acts on a first lever. The next lesson shows a second one, independent of this one: the order of words within a single message.