Why might OpenAI offer multiple model variants within the same generation?

One generation is usually three or four models, and the split is deliberate: the same family, tuned to different points on the curve where speed, cost and capability trade against each other.

Read the catalogue and the pattern is explicit. The top variant is described as "the frontier model in the GPT-5.6 family" and is priced to match. The middle one is "designed for workloads that balance intelligence and cost". The smallest is "designed for cost-sensitive, high-volume workloads", and it runs at roughly a twenty-fifth of the frontier price per million input tokens.

The names change every generation but the tiering does not — OpenAI's own page notes that a variant "roughly corresponds to the mini model tier used in earlier GPT-5 families".

This matters because most real systems use more than one. A classifier that fires on every request has completely different economics from a planning step that runs once an hour. Reaching for the frontier model everywhere is the expensive mistake; reaching for the cheapest everywhere is the unreliable one. The variants exist so that choice does not have to be made globally.