Reading a model’s name

The name is a specification

claude-3-5-sonnet-20241022. gpt-4o-mini. gemini-2.5-flash. llama-3.1-70b-instruct. qwen3-8b. deepseek-r1. These look like noise until you know the parts, and then each one tells you roughly what you are dealing with before you send a single message.

Almost every name carries four things, in some order.

The family. Who made it and which lineage it belongs to. GPT, Claude, Gemini, Llama, Mistral, Qwen, DeepSeek, Gemma, Phi.

The version. A number that goes up: 3, 3.5, 4, 4.1. A bigger number from the same maker is usually a better model, and this is the one part of the name you can mostly trust.

The size or speed label. This is the part people miss. mini, nano, flash, lite, small, haiku mean the cheap fast one. pro, opus, large, ultra mean the expensive slow one. sonnet, medium, and unlabelled names usually sit in the middle. Open models put the parameter count in directly: 8b is eight billion parameters, 70b is seventy.

The build or date. 20241022 is a snapshot from 22 October 2024. -preview, -exp, -latest are channels rather than dates, and they behave differently, which the last lesson of this module is about.

Two extra words appear on downloadable models. instruct or chat means it has been trained to follow instructions and hold a conversation: this is the one you want. A base model with no such suffix continues text rather than answering it; ask it a question and it may write three more questions. People download the base model by mistake, conclude the model is broken, and give up.

Version numbers are not comparable across companies

Gemini 2.5 is not “worse than” GPT-4 because 2.5 is less than 4. The numbers are internal to each company and reset whenever marketing decides. OpenAI has shipped gpt-4o and o4-mini as different products in the same period; one is a general model, the other from a reasoning line. Read the family first, the number second, and never compare two makers’ version numbers directly.

Finding out what is actually answering you

In a chat product, look in three places, in this order:

1. The model picker at the top of the conversation. It names a product tier, not always the exact model.
2. Settings, then the account or model page.
3. The small print under an individual answer: some products name the model there, especially after a fallback.

If it says something like ”Fast”, ”Smart”, or ”Auto”, you have been given a routing label rather than a model name. That is a real answer too: it means the product is choosing for you and may choose differently tomorrow.

Through an API you always get the exact string, because you have to type it:

curl https://api.example.com/v1/chat/completions \
  -H "Authorization: Bearer $KEY" \
  -d '{"model": "gpt-4o-mini-2024-07-18", "messages": [...]}'

Why the date on the end matters

Two names can point at the same family, the same version and the same size, and behave differently:

• claude-sonnet-4-5: an alias. It resolves to that model’s dated snapshot. Some providers’ aliases move to a newer build without notice: OpenAI’s gpt-4o has pointed at more than one dated snapshot.
• claude-sonnet-4-5-20250929: a pinned snapshot. Same weights every day until it is retired (this one on 30 November 2026).

An alias is convenient and it is also how your carefully tuned workflow changes behaviour on a Tuesday morning without anybody deploying anything. If you are running something that matters, pin the date. If you are chatting, you cannot pin anything, so keep the possibility in mind when a tool seems to have changed personality.

Snapshots do not live forever. Each provider publishes a retirement schedule. OpenAI promises at least six months’ notice for generally available models. Anthropic lists a date before which each model will not be retired (about a year after its release) and then gives at least 60 days’ notice, which in practice has been about two months. After retirement the name stops working and your script returns an error rather than a worse answer. That error is a kindness. The alias would have silently degraded instead.

The one thing a name never tells you

A model name says nothing about the product wrapped around it. gpt-4o inside a chat app with web search, file upload and memory is a different experience from gpt-4o inside a badly built browser extension that sends three sentences of context. The name identifies the engine. It does not identify the car.

So the habit is: find the exact string, write it down with today’s date next to your judgement of the tool, and remember that half of what you liked or disliked probably came from a layer the name does not cover.