Claude, OpenAI and Gemini Model Names Explained: What Opus, Astra and Carbon Mean

A2Agent Team ยท 2026-11-16T00:00:00Z

Read enough AI release notes this year and you'll notice that the three biggest US labs have mostly stopped naming models like software. Plenty of companies still use Pro, Max and Flash, but Google now ships models named after chemical elements, OpenAI after the sun and moon, and Anthropic after poems and stories.

Each scheme says something about how the company wants you to think about its models. Some of the names come with an explanation from the company itself. Others don't, and any meaning read into them is guesswork. We'll say which is which as we go.

Google: the periodic table

Gemini is the one Google name with a confirmed origin. It's Latin for "twins," and Jeff Dean has said it marks the merger of Google Brain and DeepMind into one team. The team also liked the reference to NASA's Gemini program, the 1960s missions between the first crewed flights and the Apollo moon landings.

The Gemini 4 models are named after elements:

Model Role Released
Gemini 4 Carbon Google's strongest model, for long-horizon coding, cybersecurity and research Announced October 28, 2026
Gemini 4 Argon The frontier default for complex engineering and knowledge work October 13, 2026
Gemini 4 Hydrogen The fast, cheaper model for everyday coding and agent workflows November 12, 2026

Hydrogen isn't a separate line. Its model card says it's "adapted from Gemini 4 Carbon."

Google has never explained why it chose elements. It's tempting to read Hydrogen, the lightest element, as the nimble model, and Carbon, the backbone of organic chemistry, as the structural workhorse. That's our reading, not Google's, though it fits well.

OpenAI: a solar system of sizes

OpenAI went the other way, up into the sky. Its tiers use Latin astronomy: Luna (moon), Terra (earth), Sol (sun) and Astra (stars). For months there was no official word on how they related. Then on October 28, 2026, OpenAI's Ted Sanders laid out the order in a Hacker News comment: Luna < Terra < Sol < Astra.

The names track model size. Sanders wrote that "much of the variation in performance is explained by just a single variable: model size," and that bigger models are more reliable, cost more per token and remember more obscure facts. He also wrote that "LLM performance is incredibly multidimensional," so the biggest model isn't automatically the best at every task.

The order matches the product lineup:

Tier Meaning What it's for
Luna Moon Cheap, fast models for simple requests and high-volume automation
Terra Earth Visual reasoning, computer use and moderately complex work
Sol Sun Serious analytical work where quality matters more than speed
Astra Stars The hardest reasoning, software engineering and professional work

OpenAI avoided Small, Medium and Large, and Basic and Premium. Each tier gets its own identity, and everyone already knows the moon is smaller than the sun, so the order is easy to remember. The names also make a product argument for OpenAI: you don't always need the biggest model, and picking the right size is the sensible choice.

Anthropic: poems, then stories

Anthropic took the most literary route. Its Claude tiers are named after forms of writing: Haiku, Sonnet and Opus. Dario Amodei explained the idea on the Lex Fridman podcast (episode 452). The company wanted poetry to tell models apart by size and ambition.

The forms map onto the tiers neatly. A haiku says something complete in a few syllables, and Claude Haiku is built to be fast, cheap and still capable. A sonnet is a longer, more structured form, and Sonnet is the balanced middle tier. Opus is Latin for "work," as in magnum opus, a creator's defining piece, and Opus was the flagship.

Amodei has also pointed out the limits: generations move fast, and a new Haiku can beat an old Opus on many tasks, so the names only describe a model's role within its own generation.

Anthropic has since added Fable and Mythos, from the Latin fabula and the Greek mythos, both meaning story. They sit above Opus as Anthropic's most powerful models, and the two names describe the same underlying model with different safeguards. Fable is the general release, with extra safeguards in areas such as cybersecurity and biology. Mythos has some of those safeguards lifted and goes only to vetted organizations, starting with cybersecurity partners in Project Glasswing. When Anthropic introduced Fable 5 and Mythos 5 in June 2026, it said the models are identical apart from their safeguards, so here the name tells you about the safety setup, not the size.

Three vocabularies

Company Theme Names What it evokes
Google Chemical elements Hydrogen, Argon, Carbon Science, research, technical rigor
OpenAI Celestial bodies Luna, Terra, Sol, Astra Scale, order, the cosmos
Anthropic Poetry and stories Haiku, Sonnet, Opus, Fable, Mythos Language, craft, the humanities

All three are solving the same problem: making a growing lineup recognizable. Google's elements sound like a research lab, which suits a company that publishes across AI, math, physics and biology. OpenAI's names are the most practical, because they tell you how to choose. Anthropic's make a technical product feel like a piece of culture, in the same voice the company uses when it talks about model behavior and safety.

The name is the first thing users see

Most people never read a model card or compare benchmark scores. They see a name in a dropdown, an API call or a headline, and guess from that what the model is for, so the name does a lot of the work of explaining the product.

A naming scheme has to be easy to remember, leave room for the next dozen releases, and roughly match what each model does. Google is betting on scientific seriousness, OpenAI on a sense of order that helps people pick the right size, and Anthropic on warmth and culture to make powerful models feel approachable. OpenAI's planets and Anthropic's poems at least hint at size, since everyone knows a haiku is shorter than an opus. Google's elements don't, and Fable and Mythos tell you about safeguards instead, so for those you still have to read the docs.

Model names on A2Agent

On A2Agent you call every model by its ID, such as deepseek-v4.1-flash-1017 or glm-5.3, through one API key with OpenAI-compatible and Anthropic-compatible endpoints. The Models page lists each model's provider, context window and price, so you can choose by what a model does and costs. New accounts get $5 in free credit.

Sources