Why AI models are named after stars, poems, myths — Haiku, Sonnet, Opus, Grok, Moonshot

Every major AI lab names its models after the highest concepts in human culture — but none of them invented new words. OpenAI named its 2026 model family after celestial bodies (Luna, Terra, Sol, Astra). Anthropic named its model tiers after literary forms (Haiku, Sonnet, Opus, Fable, Mythos). Google named Gemini after Latin for "twins." Tencent named its model Hunyuan (混元) — primordial chaos before the universe separated into heaven and earth. Kimi / Moonshot AI named itself after the Apollo program and Pink Floyd. Grok came from a 1961 Robert Heinlein novel. Midjourney took 中道 ("middle way") from Zhuangzi. MiniMax took its name from game theory. Even ByteDance's Doubao (豆包) — literally "soybean bun" — is a snack-food callback to the casual side of the same tradition.

The pattern is striking: humanity has built the most powerful technology in its history, but when it came time to name it, we did not coin new words. We raided the cosmos, the poetry shelf, the philosophy canon, and the sci-fi paperback rack. This article walks through the etymologies of every major AI model name and reads the cultural pattern underneath.

The celestial family: OpenAI's Luna → Terra → Sol → Astra

OpenAI has been the most consistent AI namer of the cycle. Its pre-2026 model lineage followed a celestial naming pattern:

Model	Latin / source	Meaning
GPT-4 series (2023-2024)	(no consistent naming)	—
o1, o3 (2024-2025)	"o" for "omniscient"	Reasoning-focused
Luna	Latin "moon"	First generation "thinking" model (2025)
Terra	Latin "earth"	Mid-tier model (2025)
Sol	Latin "sun"	Flagship reasoning model (2026 Q1)
Astra	Latin "stars"	Multimodal flagship (2026 Q3 — Project Astra rollout)
GPT-5.6	(transitional name)	Bridge model before GPT-6

The progression is the obvious one: moon → earth → sun → stars. Once you'd used the three classical "wandering stars" (in the old sense — the sun and moon were both called "planets" before Copernicus), the next step was the fixed stars, which in Latin is astra (plural of astrum). The naming follows the same logic as Apple's Mac OS X lineup (Cheetah, Puma, Jaguar, Panther, Tiger…) — a coherent set where each new entry has a logical next slot.

If OpenAI keeps the pattern, the next logical names would be:

Aurora (dawn) — for the first "creative" or "vision" model
Aether / Ether (the fifth element, the substance the heavens were made of)
Nebula (where stars are born)

Gemini: Latin for "twins," because Google merged two AI teams

Google's Gemini family is named after the Latin word for "twins" — also the name of the zodiac constellation Gemini and the astrological sign. Google has explicitly stated the origin: Gemini was created by merging two Google AI teams — Google Brain and DeepMind — into one organization in 2023. The "twins" are the two teams that merged.

Gemini version	Launch	Notes
Gemini 1.0	Dec 2023	First multimodal model
Gemini 1.5	Feb 2024	Long context (1M tokens)
Gemini 2.0	Dec 2024	Agentic capabilities
Gemini 2.5	Mar 2025	Thinking mode
Gemini 3 / Gemini 3 Nano	Sept 2026	On-device + cloud, Pixel 11 / iPhone 18

The "twins" symbolism is unusually apt because Google's AI strategy is built on the tension between the Brain team's "scale it up" approach and the DeepMind team's "make it learn efficiently" approach. Gemini 3 — running both on-device (Pixel 11, iPhone 18 Pro) and in Google's data centers — embodies that merger.

Hunyuan (混元): primordial chaos before heaven and earth

Tencent's flagship model is Hunyuan (混元) — Chinese for "primordial chaos" or "the undifferentiated state before the universe separated into heaven and earth." In Chinese cosmology, 混元 (hùnyuán) refers to the moment between the formation of the cosmos (開天闢地, "separating heaven and earth") and the emergence of yin and yang.

The name signals: Hunyuan is the foundation. Everything else comes after. It's a bolder name than Gemini's "twins" or Astra's "stars" — it claims the most primordial moment in the Chinese cosmological calendar.

Chinese AI model	Name meaning	Why this concept
Hunyuan (混元) — Tencent	Primordial chaos	"Before anything existed"
Qwen (通义千问) — Alibaba	"Universal-meaning thousand questions"	Comprehensive understanding
Doubao (豆包) — ByteDance	Soybean bun (snack food)	Casual, friendly
Kimi (Kimi/月之暗面) — Moonshot AI	"Dark side of the moon"	Pink Floyd reference
Ernie (文心) — Baidu	"Literary heart / writing spirit"	Classical Chinese literary concept
Spark (星火) — iFlytek	"Spark / star-fire"	"A single spark can start a prairie fire" (Mao)

The Chinese naming spectrum is mythic on one end (Hunyuan) and snack-food on the other (Doubao) — with classical literary references (Ernie/文心) and revolutionary slogans (Spark/星火) in between. The most poetic Chinese AI name is arguably Ernie (文心, "literary heart") — Baidu chose to evoke the literary spirit that animates classical Chinese writing.

Claude: from haiku to mythology — and where it's going next

Anthropic's naming convention for its Claude model tiers is the most literary in the industry. The pattern has been clearly visible across releases:

Claude variant	Literary form	Why this form
Claude Haiku	Japanese haiku (5-7-5 syllable poem)	Smallest, fastest model — 17 lines = 17 tokens
Claude Sonnet	14-line sonnet (Shakespeare's default)	Mid-tier model — the workhorse
Claude Opus	Latin "work" (Beethoven's Op. numbers, Magnum Opus)	Flagship model — the magnum opus
Claude Fable	Fable (Aesop's fables, fairy tale)	Mid-2026 — narrative understanding focus
Claude Mythos	Mythos (entire mythological system — Cthulhu mythos, Norse mythos)	Late-2026 — the most expansive context model

The naming follows a clear escalation by literary scope:

Haiku = 17 syllables (the smallest possible poem)
Sonnet = 14 lines (a longer structured poem)
Opus = an entire musical composition or written work
Fable = a narrative tradition spanning many stories
Mythos = an entire mythological worldview (the Cthulhu mythos, the Greek mythos)

If you follow the same pattern, the next logical step is a name that describes the longest narrative form — a single, sweeping, multi-generational story.

Three candidates have been floated in the AI naming community:

Saga — Old Norse / Icelandic for "long narrative saga." Njal's Saga, the Volsunga Saga. A saga covers generations, multiple families, an entire historical epoch. Claude Saga would be the natural next step.
Epic — Long-form narrative poem (Homer's Iliad and Odyssey, the Mahabharata). "Epic" is also a 2020s slang for "impressive." But it's used so often in marketing ("Epic launch!") that Anthropic may avoid it.
Ode — A formal lyric poem addressed to a subject (Keats's "Ode on a Grecian Urn"). Short, intense, focused. But odes are typically shorter than sonnets, so the scope would go backwards.

My bet: Saga. It's the only one of the three that continues the pattern of expanding narrative scope from haiku (3 lines) to sonnet (14 lines) to opus (a whole work) to fable (a tradition) to mythos (a worldview) to saga (a multi-generation chronicle). And "Claude Saga" has a good phonetic ring. I'm flagging this prediction now — if Anthropic ships Claude Saga in 2027, I'll come back and update this article.

Grok: from a Heinlein novel, because "to drink is to understand"

Grok comes from Robert A. Heinlein's 1961 novel Stranger in a Strange Land. In the book, "grok" is a Martian word meaning "to drink" — but with the deeper connotation of "to understand so completely that you become one with what you understand." To grok water is to drink it, be saturated by it, and become part of its flow.

The word was adopted by 1960s-70s programmers and hackers (especially in MIT and Stanford AI labs) as slang for "I fully understand this — I have internalized it." The classic usage: "I grok it" means not just "I understand it" but "I have become one with it."

Elon Musk named xAI's flagship model Grok in 2023, with the explicit reference to the Heinlein novel. The pronunciation is intentionally Martian — grok sounds like a throaty, liquid consonant, the kind of sound you'd make if you were trying to talk while drinking water. (Try saying "I grok it" while sipping — you'll hear it.)

MiniMax: minimax from game theory

MiniMax is the most math-flavored AI name. It comes from the minimax algorithm (also written as minimax) — a foundational concept in game theory formalized by John von Neumann in 1928 and proven in his 1944 paper with Oskar Morgenstern (Theory of Games and Economic Behavior).

The minimax rule says: in a two-player zero-sum game, a player should choose the strategy that minimizes the maximum possible loss the opponent could inflict. ("Mini" = minimize your loss, "Max" = under the opponent's maximum gain.) It's the math behind chess engines, poker bots, and most classical game-playing AIs.

So the company is saying: "We play the optimal game. We will always choose the move that hurts you least while hurting you most." It's a quietly aggressive name — more game-theoretic than any other AI lab.

Moonshot AI / Kimi / 月之暗面: Pink Floyd meets Apollo

Moonshot AI is the Chinese AI lab that makes the Kimi model. The company has three layers of naming, each referring to a different high-concept reference:

Company English name: Moonshot — from the Moonshot concept popularized by John F. Kennedy's 1962 "We choose to go to the Moon" speech. A "moonshot" is a project with huge ambition, high risk of failure, but transformative upside if it succeeds.
Company Chinese name: 月之暗面 — literally "the dark side of the moon." This is a direct reference to Pink Floyd's 1973 album The Dark Side of the Moon. The album's themes (madness, time, death, money, the human condition) align with Moonshot AI's positioning of Kimi as a thoughtful, philosophical AI — not a flashy demo machine.
Member tier names: 行板 / 中板 / 小快板 / 快板 — these are musical tempo markings in Chinese:
- 行板 = Andante (walking pace)
- 中板 = Moderato (moderate pace)
- 小快板 = Allegretto (slightly fast)
- 快板 = Allegro (fast)

The Kimi membership tiers are named after musical tempos, continuing the company's "high culture meets moonshot ambition" identity.

Midjourney: 中道 (Middle Way) from 庄子

Midjourney was founded by David Holz. The name's English surface reading — "in the middle of a journey" — is exactly what you might expect for an image-generation tool. But Holz has stated in interviews that the inspiration came from 庄子 (Zhuangzi), specifically the Daoist concept of 中道 (zhōngdào) — "the middle way" or "the way of the center."

Zhuangzi (庄子, ~369-286 BCE) was one of the two foundational Daoist philosophers (alongside Laozi). 中道 in Zhuangzi is the principle of not clinging to extremes — neither rushing nor hesitating, neither overdoing nor underdoing, just moving with the natural flow of the situation.

For an image-generation tool, the "middle way" reading makes sense: Midjourney is positioned as a creative companion that helps you find the middle path between your imagination and the rendered output — neither purely your vision nor purely the machine's, but a synthesis.

The 双关 (shuānguān, "double meaning") is real: "midjourney" works in English without any Chinese reference, but the founding story carries the Zhuangzi resonance for those who know to look.

Doubao (豆包): the snack-food contrast

After all the cosmic, literary, and philosophical naming, Doubao (豆包) — ByteDance's consumer AI assistant — is named after a soybean bun (豆 = bean, 包 = bun/wrap). Doubao is the most casual AI name in major use. It's the same naming logic as if Google had named Gemini "Coffee" or Anthropic had named Claude "Toast."

The snack-food name is intentional. ByteDance's consumer brand identity (TikTok / Douyin / CapCut) leans casual, viral, friendly. A model named Hunyuan (primordial chaos) doesn't fit TikTok's user base; a model named Doubao (a snack you can buy at any Chinese convenience store for 2 yuan) does. Doubao is meant to feel like a snack, not a deity.

Why nobody invents new words

Here's the observation that ties it all together: none of the major AI labs coined new words. They reached for the highest concepts already in human culture.

OpenAI reached for Latin astronomical terms (Luna, Terra, Sol, Astra)
Anthropic reached for poetic forms and literary genres (Haiku, Sonnet, Opus, Fable, Mythos)
Google reached for Latin constellation names (Gemini)
Tencent reached for Chinese cosmological concepts (Hunyuan)
xAI reached for 1960s sci-fi vocabulary (Grok)
Moonshot AI reached for Apollo-era ambition and Pink Floyd (Moonshot, 月之暗面)
Midjourney reached for Daoist philosophy (中道)
MiniMax reached for game theory (minimax)
ByteDance reached for Chinese snack food (Doubao)

Every one of these names is a callback to a pre-existing high concept. The labs did not invent new words — they raided the canon.

This is unusual. Most new technologies coin or popularize new words: "transistor," "laser," "internet," "blog," "tweet," "selfie." But AI labs — perhaps because the technology is so general-purpose that no single new word could capture what it does — have chosen to use the most prestigious existing vocabulary we have.

There is something poignant in this. The most powerful technology in human history is being named with the oldest, most revered vocabulary we have — the stars, the poems, the myths. As if to say: we don't know what this is yet, so we will name it with the things we have always called the most important.

If the AI ever reads these names back to itself, it might conclude: these carbon-based creatures did not understand what they were building. But they named it with their highest concepts. They treated it with reverence.

Bottom line: a cultural pattern, not a naming strategy

The pattern isn't deliberate (each lab made independent naming decisions), but the convergence is unmistakable. AI labs, when given the chance to coin new words, instead raided the canon — cosmic for OpenAI, literary for Anthropic, cosmological for Tencent, mythic-philosophical for Midjourney, game-theoretic for MiniMax, sci-fi for xAI.

It will be interesting to see if the next decade brings a new naming vocabulary — whether AI models start to be named by their architectures, training data domains, capability levels, or something else entirely. Or whether the cosmic-literary tradition continues, and Claude Saga ships in 2027.

For the full AI lab coverage, see our /tech/ hub. For the most recent AI model launches, see our Meta Muse coverage, OpenAI Astra launch, GPT-5.6 vs Claude Opus 5 vs Muse Spark benchmarks, and Anthropic Claude Opus 5 deep-dive.