AI Model Names Are Getting Out of Hand: Tracing GPT, Gemini, and Claude

AI Model Names Are Getting Out of Hand: Tracing GPT, Gemini, and Claude

Generative AI made major strides in 2026.
From coding and research to preparing documents, there is more work we can delegate to these models than ever before.

Yet as their capabilities grow, so does the number of models to choose from. Figuring out which one to use has become surprisingly complicated.

Google recently introduced Gemini 4 Argon, followed by Carbon.
I wondered whether Carbon was the new Flash equivalent, but apparently not.
Then came Hydrogen and Neon, adding two more element names to the lineup.

OpenAI has GPT-5.1-Codex-Max and GPT-5.3-Codex-Spark, while the GPT-6 family uses Astra, Sol, Terra, and Luna.
Anthropic has added Mythos and Fable alongside its familiar Opus, Sonnet, and Haiku series.

Better models are welcome. But their names are making it harder, not easier, to understand what each one is for.

I lined up the generations from all three companies to see how their naming schemes reached this point.

OpenAI: version numbers, specializations, nicknames, and Max

OpenAI's GPT-4 era leaned on numbers and size labels such as mini.
The GPT-5 generation added more Codex variants. GPT-5.6 introduced Sol, Terra, and Luna, followed by Astra in GPT-6.

GenerationGeneral / standardPremium / specializedLightweight / fast
4GPT-4——
4.1GPT-4.1—GPT-4.1 mini
4.5GPT-4.5 Preview——
4oGPT-4o—GPT-4o mini
5GPT-5GPT-5 proGPT-5 mini
5—GPT-5-CodexGPT-5 nano
5.1GPT-5.1GPT-5.1-CodexGPT-5.1-Codex-Mini
5.1—GPT-5.1-Codex-Max—
5.2GPT-5.2GPT-5.2-Codex—
5.3GPT-5.3 ChatGPT-5.3-CodexGPT-5.3-Codex-Spark
5.4GPT-5.4—GPT-5.4 nano
5.6GPT-5.6 TerraGPT-5.6 SolGPT-5.6 Luna
6GPT-6 SolGPT-6 AstraGPT-6 Luna
6.1GPT-6.1 TerraGPT-6.1 Sol—
6.2GPT-6.2 SolGPT-6.2 Astra—

This is a selection of general-purpose and developer-oriented models, not an exhaustive catalog. Image, audio, Realtime, and some variants are omitted.
Where a generation contains multiple branches, they occupy separate rows. A dash means no model is listed in that particular cell, not that no model of that version exists.

In GPT-5.1-Codex-Max, "Max" is part of the model's name.
By contrast, the API's `reasoning effort: max` setting for GPT-6.2 Astra controls reasoning effort; it does not identify a separate model named "GPT-6.2 Astra Max."
The distinction is easy to miss.

The table does not yet list a GPT-6.2 Terra.
GPT-6.1 Terra arrived after Sol and Luna, though, so a 6.2 version could still follow.
And the role of Sol has shifted: in the 5.6 lineup it occupied a higher tier, whereas the 6.2 version is positioned for more everyday use.
A familiar nickname does not guarantee the same role in the next generation.

Sources: OpenAI's complete API model directory, current model catalog, and the GPT-5.3-Codex-Spark announcement.

Google: from Pro and Flash to chemical elements

Gemini long used the relatively straightforward Pro, Flash, and Flash-Lite labels.
Flash then received a series of smaller version increments in Gemini 3, only for Gemini 4 to introduce Argon, Carbon, Hydrogen, and Neon instead.

GenerationHigh capability / reasoningFast / generalLightweight
1.5ProFlash—
2.0Pro (Experimental)FlashFlash-Lite
2.5ProFlashFlash-Lite
3ProFlash—
3.1Pro—Flash-Lite
3.5——Flash-Lite
3.6—Flash—
3.7—Flash—
3.8—Flash—

Gemini 4 no longer fits neatly into those columns.

ModelMain role or characteristic
ArgonFrontier reasoning and coding work
CarbonDifficult problems and long-running agent tasks
NeonStrong reasoning with better speed and cost efficiency
HydrogenFast, lower-cost everyday tasks

These descriptions summarize Google's stated use cases; they are not exact successor mappings to the earlier lineup.
Google also describes Hydrogen as a migration option for Gemini 3.8 Flash users.
That relationship is useful, but the name itself tells you almost nothing about it.

Chemical element names might be memorable. Their atomic numbers are Carbon (6), Neon (10), Argon (18), and Hydrogen (1), however, and they do not map to the models' performance rankings.
I thought the periodic table was something I needed for chemistry class. Apparently I also need it to choose an AI model.

Sources: Gemini API model catalog, Gemini 4 Hydrogen announcement, and Gemini 4 Neon announcement.

Anthropic: Opus, Sonnet, and Haiku were at least consistent

Anthropic used Opus, Sonnet, and Haiku across multiple generations.
Opus means a work or composition, a sonnet is a structured poem, and haiku is a Japanese poetic form.
Those meanings do not inherently tell you which model is the most capable.
What helps is that Anthropic consistently assigned the names to high-end, balanced, and lightweight roles.

Not every tier gets every minor version, though, and Claude 5 introduced Mythos and Fable as additional series.

GenerationSpecialized / demandingOpusSonnetHaiku
3—Opus 3Sonnet 3Haiku 3
3.5——Sonnet 3.5Haiku 3.5
3.7——Sonnet 3.7—
4—Opus 4Sonnet 4—
4.1—Opus 4.1——
4.5—Opus 4.5Sonnet 4.5Haiku 4.5
4.6—Opus 4.6Sonnet 4.6—
4.7—Opus 4.7——
4.8—Opus 4.8——
5Mythos 5Opus 5Sonnet 5—
5Fable 5———
5.1Mythos 5.1———
5.1Fable 5.1———
5.2Mythos 5.2———
5.2Fable 5.2———
5.5—Opus 5.5Sonnet 5.5Haiku 5.5

Mythos and Fable can share version numbers without sharing availability or positioning.
Fable 5.2 is presented for demanding reasoning and long-horizon agent work, while access to the Mythos series is restricted.
Treating all these names as one simple hierarchy above Opus would be misleading.

Even so, Anthropic's table may be the easiest of the three to understand because its core tier names have remained relatively stable.
The missing cells, irregular updates, and new branches still make the full family harder to track.

Sources: Claude model overview, deprecation history, and Fable and Mythos 5.2 announcement.

Other vendors are hardly immune

Alibaba's Qwen combines labels such as Max, Plus, Flash, and Coder with parameter-count names such as `397B-A17B`.
The latter can convey useful architectural information, but matching a hosted product name to an underlying model is not always obvious.

DeepSeek has used V and R families alongside newer Pro and Flash labels.
Mistral AI has Large and Small as well as Codestral, Devstral, Magistral, and Ministral.
Z.ai's GLM lineup mixes version numbers with Flash variants.

Each has its own logic. As model uses, architectures, and delivery methods multiply, however, expecting one name to communicate everything becomes increasingly unrealistic.

Sources: Qwen documentation, DeepSeek API documentation, Mistral model catalog, and Z.ai documentation.

The real problem is mixing different kinds of information

The problem is not that vendors are creative. It is that several distinct dimensions get crammed into or presented alongside model names:

  • Generation: numbers such as GPT-5.3, Gemini 3.8, and Claude 4.8.
  • Tier or task: Pro, Flash, Opus, Haiku, and Codex.
  • Nickname: Astra, Sol, Argon, Carbon, and Neon.
  • Runtime setting: Thinking, high, xhigh, and max may describe how a model runs rather than a distinct model.
  • Availability: Preview, Experimental, deprecated, or restricted access.

A product's display name may also differ from the API model identifier.
Upgrading can change performance, pricing, latency, and limits.
When integrating a model, developers need to know whether an ID is immutable or an alias, which model it succeeds, and whether existing test results still apply.
Too much of that information has to be hunted down separately.

Keep the names, standardize the explanation

Vendors do not have to adopt identical branding.
Not every model fits into a three-tier Pro / Flash / Lite hierarchy either.

But a common way of publishing the following would help:

  • Immutable model identifiers versus changeable aliases.
  • Versions and explicit predecessor / successor relationships.
  • Intended workloads and differences in capability, cost, and speed.
  • Model variants versus reasoning settings.
  • Release states such as Preview, Stable, Deprecated, and Retired.

An industry-wide naming and versioning guideline would not be unreasonable.
Vendors could retain their creative names while making the relationships between models much easier to understand.

Personally, I miss the simplicity of the GPT-5 era, when base, Pro, and mini largely told you what you needed to know.

I welcome better models. I just wish picking one did not start with decoding its name.

This overview draws on official documentation available in November 2026 and covers selected major models, not every release. Check each vendor's current catalog before deployment.

Related Posts

Comments

Leave a comment

Your email address will not be published. Required fields are marked.

Prove your humanity: 2   +   5   =