posts

A Tree of AI Model Names

Model names are weird. What started with GPT-2 and GPT-3 is now a hodgepodge of decimals (GPT-3.5, Sonnet 3.7, Opus 4.6, Grok 4.1) skipped version numbers (o2 where art thou?) and bolted-on descriptors (what is claude-opus-4-5-20251101-thinking-32k??)

It'd help if we could visualize this.

Let's get this out on a tree:

Nice.

The link to the yaml file for the model names is available here Contributions are welcome!

We started with GPT-2 and GPT-3. Now Phi-4-mini-reasoning and Qwen3-235B-A22B and Llama-3.1-Nemotron-70B and R1-1776 are all real model IDs that real people are expected to compare.

It's going to keep getting worse. Every company is running multiple product lines with overlapping version numbers and inconsistent tier names. Fruit from git branches that diverged six months ago are thrown away and harvested at the same time.

The Greatest Hits

OpenAI
The o2 Problem
Reasoning models go o1 → o3. They skipped o2 because O2 is a European telecom brand.
OpenAI
Version Time Travel
GPT-4.1 was released in April 2025. GPT-5 came in August 2025. A model called 4.1 came after 5. It's a separate product line.
OpenAI
There Is No o4
o4-mini exists. Regular o4 does not. They released the mini without the full version.
Google
The Great Tier Rename
PaLM 2 used animal sizes: Gecko, Otter, Bison, Unicorn. Gemini switched to Nano, Pro, Ultra, Flash. The animals were never spoken of again.
Google
Nano Banana
A model called "nano-banana" appeared on LMArena benchmarks. Sundar Pichai tweeted 🍌🍌🍌. It turned out to be Gemini 2.5 Flash Image
Mistral AI
The -stral Cinematic Universe
Every product must rhyme: Code → Codestral. Vision → Pixtral. Math → Mathstral. Small → Ministral. Reasoning → Magistral.
Meta
The Case Change
"LLaMA" stood for Large Language Model Meta AI. In version 2, it became "Llama". Just a regular word.
DeepSeek
R1-Zero
Named like a German sedan: R1-Zero, then R1, then R1-Distill. Outperformed OpenAI's o1 at 95% lower cost. The naming was the least disruptive thing about it.
Anthropic
Version Hopscotch
Versions shipped: 3, 3.5, 3.7, 4, 4.5, 4.6. There is no 3.6. Claude 3.5 Opus was announced but never shipped. Haiku 4 doesn't exist. It jumped from 3.5 to 4.5.
Apple
Radical Anti-Naming
Apple called their model "Apple Foundation Models." The two variants are AFM-on-device and AFM-server. That's it. They stopped there.
Microsoft
Phi-4-mini-reasoning
The full model name is Phi-4-mini-reasoning. Model family + version + size tier + capability. Four concepts in one hyphenated name. Also: Phi-4-reasoning-plus.
OpenAI
Codex: Back From the Dead
Codex was discontinued in March 2023. In 2025, the name reappeared as GPT-5.2-Codex. They brought it back.

Last updated February 2026. All names are real. None of this is satire.