What the model numbers actually mean

The naming conventions are confusing by accident, not design

AI companies didn’t set out to confuse you with their model names. The naming systems evolved as products got more complex — and they haven’t always been consistent.

But there are patterns. Once you see them, the names become readable.

The basic structure: family, version, size

Most AI model names follow a rough structure: the model family (what kind of model it is), a version number (how recent it is), and sometimes a size indicator (how capable it is within that family).

Claude: Anthropic names models by capability tier — Haiku (fastest, cheapest), Sonnet (balanced), Opus (most capable). Plus a version number. So “Claude Sonnet 4” is the fourth generation Sonnet model — fast and capable, not the most expensive.

GPT: OpenAI uses numbers for generations (GPT-4 is newer than GPT-3) and letters for variants. “GPT-4o” — the “o” stands for “omni,” meaning it handles text, images, and audio natively. “GPT-4o mini” is a smaller, faster, cheaper version.

Gemini: Google uses names (Nano, Flash, Pro, Ultra) for size tiers and numbers for generations. “Gemini 1.5 Pro” is the Pro tier of the 1.5 generation.

What the size tiers mean in practice

Every major AI company now offers their models in multiple sizes. The pattern is consistent across companies:

Small/fast/cheap models (Haiku, Mini, Flash, Nano): Lower capability but much faster responses and lower cost. Good for simple tasks, high-volume applications, or situations where speed matters more than quality.

Mid-tier models (Sonnet, standard GPT-4o, Pro): The sweet spot for most tasks. Strong capability without the cost and latency of the largest models. This is what most people use most of the time.

Flagship models (Opus, GPT-4, Ultra): Maximum capability. Slower and more expensive. Worth it for complex reasoning, difficult writing tasks, or anything where quality matters more than speed.

How to choose in practice

For everyday use through a chat interface — you don’t need to choose. The service picks a sensible default. When you’re using the API or building something, the choice matters because it affects cost and performance.

The simple rule: start with the mid-tier model. If it’s not good enough for your task, move up. If cost is a concern and quality is sufficient, move down.

Why versions matter

A higher version number generally means the model has been trained on more data, with better techniques, producing better results. GPT-4 is better than GPT-3. Claude 3 is better than Claude 2. This doesn’t always mean you need the newest model — older models often handle common tasks perfectly well and cost less — but it does mean that if you’re getting bad results, checking whether you’re using a current model is worth doing.