AI Model Names and Sizes
Level 2 · Names, sizes and licenses
How to read a model name: the family, the size tier and the version. What gpt-oss-20b, Mixtral 8x7B and o4-mini each tell you.
Course contentsLesson 6 of 13
Level 1 · The AI labs
Level 2 · Names, sizes and licenses
Level 3 · Picture, video and speech models
Level 4 · Where models live and run
After this lesson you can
- Read the family, size and version in a model name
- Work out what a size tier does to the price
- Read a parameter count from a name
12 mincovers 5 glossary entries
Worth reading first: OpenAI, Anthropic and Google
How to read a model's name
A model's name packs in three facts: the family, the size and the version. In Claude Opus 5.5, Claude is the family, Claude Opus is the size tier and 5.5 is the version. Anthropic's tiers run from Claude Haiku, the fastest and cheapest, up to Claude Fable, the strongest. Most labs name their models this way.
The tier changes the price a lot. Models are priced per million tokens, the small chunks of text they read and write. In November 2026, Claude Haiku 5.5 started at $0.10 per million input tokens, and Claude Fable 5.2 cost $8. Google's Gemini comes as Gemini Flash for speed and Gemini Pro for strength. A bigger tier is slower and dearer, and often not needed.
Open models often put the size right in the name, as a count of parameters. The B means billion. gpt-oss-20b has about 21 billion parameters, and gpt-oss-120b about 117 billion. The count tells you how much memory the model needs, and so which machine can run it.
Some names hide a twist. Mixtral 8x7B, from December 2023, has eight groups of experts in each layer, and a router uses two of them for each token. So it holds about 47 billion parameters, not 56, and uses about 13 billion per token. It runs at the speed of a much smaller model.
OpenAI breaks the pattern with its o-series. o1-preview, in September 2024, was its first model built to think before it answers. Later came o1, o3-mini, o3 and o4-mini. The o line is not part of GPT. A higher number is not always a bigger model, and mini means a small, cheap one.
The version moves on often, so labs also give each release a fixed ID, called a snapshot. A short name may move to a newer model without warning. A snapshot stays the same until the lab retires it, on a date it announces ahead. Code that must behave the same next month should name the snapshot.
People read a higher number as a better model for every job. A new small model can beat an old big one, and a model built to think can be slow and costly on easy tasks. The name tells you the family, the size and the age. Only a test on your own task tells you which is best.
Four model names, taken apart
| Name | Family | Size | Version |
|---|---|---|---|
| Claude Opus 5.5 | Claude | Opus tier | 5.5 |
| gpt-oss-20b | gpt-oss | About 21 billion parameters | First release |
| Mixtral 8x7B | Mixtral | Eight experts, about 47 billion in all | First release |
| o4-mini | OpenAI o-series | Mini, the small size | 4 |
Worked example
Jack picks a tier for his help desk
Jack's help desk sorts 20,000 emails a day, about 500 tokens each. That is 10 million input tokens a day.
At $0.10 per million, Claude Haiku costs him about $1 a day. Claude Opus, at $4 per million, would cost $40. He tests both on 200 emails, both sort every one right, and he keeps Claude Haiku.
The words for this lesson5
Check yourself
What does the number in gpt-oss-20b tell you?
Finished this lesson?
Marking it complete also marks its 5 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.