The Pour Over logo
TPO ExplainsThe Pour Over

What's in an AI model name?

GPT-4o. Claude 3.5 Sonnet. Llama 405B. Gemini Flash. Grok. The people building AI apparently pulled their naming conventions out of five different hats. Here's a plain-English decoder for the labels you keep bumping into — plus the surprisingly nerdy stories behind them.

Read one label
Claude3.5Sonnet
Claude
The family. Basically, who made it.
3.5
The version. A half-step newer than 3.
Sonnet
The size. The comfy middle option.
Most names follow the same three-part combo: family, version, size. The rest is mostly marketing.
Meet the families
GPTOpenAI

GPT is short for Generative Pre-trained Transformer, a real mouthful that describes the tech (not a cute mascot, sorry). The number is the generation: bigger means newer and smarter.

  • 4o: the “o” stands for omni. One model that juggles text, images, and audio.
  • o1, o3:a separate “reasoning” line that actually thinks before it blurts. And no, you didn't miss o2 — they skipped it to dodge a clash with a UK phone company.
  • mini / nano: the pocket-sized versions. Cheaper, faster, and lighter than the full model.
ClaudeAnthropic

Named after Claude Shannon, the genius who basically invented information theory. The sizes borrow names from writing formats, shortest to longest:

  • Haiku: small, quick, and to the point.
  • Sonnet: the reliable everyday workhorse.
  • Opus: the big, brainy one.

The number (3, 3.5, 4) is just the version.

GeminiGoogle

Gemini means “the twins,” a nod to the moment Google's two AI teams (Brain and DeepMind) became one. Bonus points for also winking at NASA's old Gemini program.

  • Ultra / Pro / Flash / Nano: sizes from heavyweight down to featherweight.
  • Gemma:the little open-source sibling anyone can download. It's Latin for “gem.”
LlamaMeta

Llama is a backronym for Large Language Model Meta AI. That it spells an adorable animal is a happy accident. Meta hands these out as “open weights,” so anyone can run them.

  • The number (2, 3, 4) is the version.
  • 8B, 70B, 405B: the “B” is billions of parameters, a rough measure of size. 405B is a giant brain; 8B is happy on your laptop.
GrokxAI

Grok comes from the 1961 sci-fi classic Stranger in a Strange Land, where “to grok” means to understand something so deeply it becomes part of you. The number is just the version.

The lingo, quickly
Parameters
The model's internal knobs. More of them (70B, 405B) usually means more brainpower.
Tokens
The little chunks of text a model reads and writes. Think a word, or a piece of one.
Context window
How much it can hold in its head at once. A page, or a whole novel.
Multimodal
Handles more than text. Images, audio, even video.
Open weights
Free to download and run yourself, like Llama and Gemma.
Reasoning model
Takes a beat to actually think through the hard stuff before answering.
Why it still feels like alphabet soup

There's no shared rulebook. Every lab names things its own way, and the labels keep shifting release to release.

  • A bigger number isn't always newer or better.
  • 4o and o4 are two different models.
  • A “mini” can still smoke last year's full-size model.
The short version: name, version, size. The rest is marketing.
Return to TPO Explains