AI Model Names Explained: What GPT, Gemini, Claude and Llama Version Numbers Actually Mean
GPT-6 Astra, Gemini 3.8 Flash, claude-sonnet-4-6 and Llama 4 Scout decoded: what generation numbers, family words, snapshot dates and Preview labels actually signal — and why numbers never compare across providers.
AI model names look like secret code, but each one is a generation number, a family word and sometimes a date or label — read those three parts, and the picker stops feeling random.
Key points
- Every provider uses the same grammar: generation number, family word, and sometimes a date or status label.
- Numbers only compare inside one provider and one family — a higher number elsewhere means nothing for you.
- Claude pins each ID to one fixed snapshot, and Google labels how settled a version is.
Every model name reads in three parts
Generation first, family second, then a job or status word.
If a model picker has ever left you frozen — three similar names, no idea which you used yesterday — the names really are dense. They pack a generation, a family and a job into a few words.
Read in order, it gets kinder: the number is the generation, the word after it is the family or tier, and anything after that tells you how settled the release is, or what it is for. Our pillar on how AI model updates work fills in the rest.
OpenAI: generation first, then a tier word
GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna decode at a glance.
OpenAI starts with the generation — GPT-6, or the point release GPT-6.1 — then adds a tier word. Its models page calls GPT-6 Astra the flagship for complex reasoning and coding, GPT-6.1 Sol a balance of intelligence and cost, and GPT-6 Luna the most efficient model for high-volume work, with GPT-6.1 Terra fast and efficient in between.
The model ID is the same name in lower case: gpt-6-astra, gpt-6.1-sol. Specialised models add a task word, such as the Cyber models for authorised security work. In our view, this is the easiest system to rank.
Google: a number, a tier word, then a status
Flash and Pro name the tier; Stable, Preview and Latest say how settled it is.
Google’s Gemini page lists Gemini 4 and Gemini 3, and numbers move at different speeds inside one generation: Gemini 3.8 Flash and Gemini 3.6 Flash are stable, while Gemini 3.1 Pro uses the endpoint gemini-3.1-pro-preview. The word after the number names the tier or task — Flash, Flash-Lite, Pro, Live for live audio, TTS for text-to-speech.
Gemini 4 changes the top tier words: Gemini 4 Carbon is called its strongest reasoning model, and Gemini 4 Argon an efficient frontier model, faster with fewer tokens per turn.
Claude: the family sits inside a pinned ID
Sonnet, Opus and Haiku are the family; older IDs end in a snapshot date.
Anthropic gives the clearest rule here: each Claude model ID identifies a pinned version, and the underlying model remains constant for the lifetime of that ID. The family word sits in the middle — claude-sonnet-4-6, claude-opus-4-7 and claude-opus-4-8 are its examples.
Older IDs include a snapshot date, such as claude-sonnet-4-5-20250929 and claude-haiku-4-5-20251001. From the 4.6 generation, IDs are dateless, major releases like Claude Sonnet 5 and Claude Opus 5 omit the minor segment, and a dateless ID is the snapshot itself — not a pointer that quietly moves.
Llama: a generation, point steps, a variant
Llama 4, then 3.1, 3.2 and 3.3, with names like Scout for variants.
Meta’s scheme is the plainest. Its developer page lists Llama 4, then the Llama 3 family including 3.1, 3.2 and 3.3 across instruction-tuned text models, and Llama 2 before that. The number is the generation; the point releases are steps inside it.
Added words name a variant or job: Llama 4 Scout is listed as natively multimodal, and Llama Guard 4 as updated protection models for Llama 4.
Point numbers, dates and suffixes show how settled a model is
A point step refines, a date fixes a snapshot, and Preview or Latest warns of change.
Point numbers work alike everywhere: GPT-6.1, Gemini 3.8 and Llama 3.3 are steps inside a generation. Dates fix a moment — in older Claude IDs, and in Google’s preview example gemini-2.5-flash-preview-09-2025. Google spells its labels out: stable points to a specific model that usually does not change; preview models will be deprecated with at least two weeks’ notice; a latest alias like gemini-flash-latest is hot-swapped each release, with two weeks’ email notice first.
Claude promises the opposite stability: weights are fixed for a given ID, though serving infrastructure — request router, safety classifiers, sampling logic — can change and cause minor differences. To check your model, see how to tell which AI model version you are using.
| Provider | Family / tier names | What number, date or suffix signals | Example pattern |
|---|---|---|---|
| OpenAI | Astra (flagship), Sol, Terra, Luna | Generation or point release; tier ranks capability against cost | GPT-6 Astra, ID gpt-6-astra |
| Carbon, Argon, Pro, Flash, Flash-Lite; Live, TTS | Step inside a tier; Stable / Preview / Latest / Experimental = how settled | Gemini 3.8 Flash; gemini-3.1-pro-preview | |
| Anthropic (Claude) | Sonnet, Opus, Haiku, Fable | Pinned version; pre-4.6 IDs end in a snapshot date, 4.6+ are dateless | claude-sonnet-4-6; claude-sonnet-4-5-20250929 |
| Meta (Llama) | Scout; Guard for protection models | Generation; point number = step inside the family | Llama 4 Scout; Llama 3.1, 3.2, 3.3 |
What this means
Learn one provider’s grammar, and treat other providers’ numbers as foreign currency.
The confusion in a model picker is not a knowledge problem. It is a design problem. You open an app to write an email or finish homework, and instead you must choose between what look like trim levels on a car you never wanted to buy. Of course you guess. Of course you then worry you guessed wrong, and paid in money or answer quality for a model you did not need.
The calming truth is that you only ever read a name in one direction: inside one provider, inside one family. A newer Sonnet number means a newer Sonnet. A Flash with a higher number than another Flash is the later Flash. Cross a border — a Gemini number against a GPT number, an Opus against a Sol — and the numbers stop meaning anything, because each company counts its own generations and moves its tiers at its own speed. Google’s catalogue proves it quietly: Flash sits at 3.8 while Pro sits at 3.1. Neither is behind. They run on different clocks.
In our view, the habit that works is boring. Pick the tier matching the job — the efficient, high-volume option for everyday sorting and short answers, the flagship when reasoning really has to hold — and stop re-deciding daily. Note the full name of any model a good workflow depends on, like a recipe that worked. If answers later feel different, that note tells you whether the model changed, or whether you are meeting the same fixed weights through slightly changed plumbing. Preview means short-notice retirement; Latest means a swap is coming; a pinned ID means neither. That turns a small daily anxiety into a ten-second decision you can explain to yourself, and it works for a student, a freelancer or anyone whose side project quietly depends on one particular model behaving the same way tomorrow as it did today.
Bottom line
Read generation, family and status separately — and never compare numbers across providers.
A model name is three facts in one coat: the generation, the family, and how fixed the release is. Learn the order for your own tools, favour stable and pinned names, and the picker stops looking like code.
Frequently Asked Questions
Cost, safety and cross-provider questions the sections above do not cover.
Does a higher version number always mean a model costs more?
No, a higher version number does not always mean a higher price. On OpenAI’s models page, the tier matters more than the step: GPT-6 Astra is listed at $10 per million input tokens, GPT-6.1 Sol at $2, and GPT-6 Luna at $0.10, so family and job set the price far more than the point release.
Is it safe to use a preview model for important work?
No, a preview model is not the safest choice for work you cannot afford to redo. Google says preview Gemini models can carry more restrictive rate limits and will be deprecated with at least two weeks’ notice, while experimental models are typically not suitable for production use. For dependable work, Google recommends a specific stable model instead.
Can you compare a Gemini number directly with a GPT or Claude number?
No, you cannot compare version numbers directly across providers. Each company numbers its own generations and moves each family on its own schedule, so Google can list Gemini 3.8 Flash as stable while Gemini 3.1 Pro is still a preview in the same catalogue. A number only gains meaning inside one provider and one family.
{“@context”:”https://schema.org”,”@type”:”FAQPage”,”mainEntity”:[{“@type”:”Question”,”name”:”Does a higher version number always mean a model costs more?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No, a higher version number does not always mean a higher price. On OpenAI’s models page, the tier matters more than the step: GPT-6 Astra is listed at $10 per million input tokens, GPT-6.1 Sol at $2, and GPT-6 Luna at $0.10, so family and job set the price far more than the point release.”}},{“@type”:”Question”,”name”:”Is it safe to use a preview model for important work?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No, a preview model is not the safest choice for work you cannot afford to redo. Google says preview Gemini models can carry more restrictive rate limits and will be deprecated with at least two weeks’ notice, while experimental models are typically not suitable for production use. For dependable work, Google recommends a specific stable model instead.”}},{“@type”:”Question”,”name”:”Can you compare a Gemini number directly with a GPT or Claude number?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”No, you cannot compare version numbers directly across providers. Each company numbers its own generations and moves each family on its own schedule, so Google can list Gemini 3.8 Flash as stable while Gemini 3.1 Pro is still a preview in the same catalogue. A number only gains meaning inside one provider and one family.”}}]}
Related Guides
Keep exploring how models are named, checked and swapped.
- AI Model Updates Explained: How ChatGPT, Gemini and Claude Get New Versions
- How to Tell Which AI Model Version You Are Using (and Why It Sometimes Changes)
- Mistakes to Avoid When a New AI Model Launches
- How to Switch AI Models Without Losing Your Chats
Sources
The official naming and model pages behind every claim above.
