Six hours after OpenAI’s launch of GPT-4.1, Sam Altman was already apologizing.
This time, it wasn’t about hallucinations or bias or Scarlett Johansson. No, it was about the model name. GPT-4.1 seemed nonsensical to many, difficult to parse from their already launched models like GPT-4o and GPT-4.5. “How about we fix our model naming by this summer and everyone gets a few more months to make fun of us (which we very much deserve) until then?” Altman wrote.
Streamers take the brunt of the internet’s name-mocking: Are you a Hulu, Tubi, or Fubo subscriber? But AI companies are just as bad, if not worse. Their model names are often incoherent and unmemorable. From Sonnet to Llama, AI companies have a branding problem of their own creation.
The history of bad AI model names
OpenAI chose a numeric naming system for their models. This streamlined their products, moving from GPT-1 to GPT-2, GPT-3, and eventually GPT-4. Now, instead of continuing their linear progression, OpenAI is launching models within the GPT-4 umbrella. First, there was GPT-4o (as in “omni”), then GPT-4.5. Now, they’ve reversed back to GPT-4.1. Users are left not knowing which models do what, or which is the most recent.
Anthropic was founded by OpenAI alums—and they have the same problem with model naming. They’ve now moved to decimals, tracking from Claude 1 to 2 to 3, before pivoting to 3.5 and 3.7. Worse, their individual models are named after literary works: Opus, Sonnet, and Haiku. Naming your models after a creative class you threaten to displace is an act of cruelty.
Google named its model Gemini because of the “dual-natured personality” of the zodiac sign. The choice is sweetly symbolic—until you realize Google is calling its chatbot two-faced. Iterations of Gemini also come in decimals, but at least they’re consistent in moving in halves, from 1.0 to 1.5 to 2.0 to 2.5. But they also have addendums like Flash, Flash-Lite, Pro, Ultra, and Nano—terms that are minimally descriptive.
But Meta is the worst among them. They chose to name their family Llama—or originally LLaMA, short for Large Language Model Meta AI, though the capitalization has been mostly given up on. It’s an overly smiley, somewhat cloying choice. No, Meta’s family of LLMs is not a petting zoo. Moreover, their individual models are named things like Scout and Maverick, because Meta seems to think their AI needs to sound like Top Gun movies.
What should an AI company look like?
A model name should have some basic tenets. It should show the company’s progress, and allude to the model’s use cases. (For how confusing GPT-4o was, the “omni” did demonstrate its value, which was that it was multimodal.) But the model names are also a prime opportunity for branding.
The AI companies knew as much when they were first launching. The initial race was between DeepMind and OpenAI, both of whom’s names reflected their positions. DeepMind was “deep,” placing their tech at a high value as they raced towards AGI. OpenAI was “open,” centering transparency for the then-nonprofit. When Anthropic came around, promising to produce more ethical AI, they centered the human in their name.
Model names can do the same. They don’t need to be zodiac signs and poems, farm animals and number splatters. These models can present their companies’ futures; they just need to pivot.