Why think in families?
Because the specific model names get swapped out every six to eight months, while the families have been the same for years. If you know each vendor's strength and philosophy, a new release is not a surprise but something you can place.
This article deliberately lists no version numbers. What is here will still be true in six months; a version number may not survive the week.
The Western closed models
These are the best known, and they follow much the same pattern: no downloadable weights, access through an API or a chat interface, and within every family a cheap tier, a middle tier and a top tier.
- OpenAI — GPT. The widest product range: chat, API, images, audio, coding. This is the most recognised brand, and typically the one a non-technical audience meets first.
- Anthropic — Claude. Strong at long, multi-step work and at coding. Prominent on the enterprise and developer side, more restrained in consumer marketing.
- Google — Gemini. The most tightly built into an existing ecosystem: search, Workspace, Android, cloud. The broadest on the multimodal side.
- xAI — Grok. Tied to the X platform, leaning heavily on real-time content.
- Meta. For years Llama was the reference for open-weight models; the company has since moved its frontier work in a closed direction, and the role of open releases in its portfolio has become less certain.
The European player
Mistral is the French — and the only genuinely significant EU — player at the frontier. It offers both closed and open-weight models, separate lines for coding, reasoning and audio, and, which for an EU buyer is often the point, an EU endpoint.
That matters less for quality than for law: in many organisations procurement's first question is not which model is better but which one you can sign a contract with at all. European AI Options works through that.
The Chinese open-weight wave
This is the part most local coverage leaves out, even though it moves half of the market measured in volume. These models ship with downloadable weights, often under permissive licences, and they run orders of magnitude cheaper.
- DeepSeek — the reference for cost efficiency, with MIT-licensed releases.
- Alibaba — Qwen — a wide range from small to top tier, with long context windows.
- Z-AI — GLM, Moonshot — Kimi, MiniMax, Tencent — all active, all with open-weight releases.
There is an American open line too — NVIDIA and OpenAI itself have both released downloadable models — but the volume today comes from the Chinese labs. Why that matters even if you never run one is the subject of Why Open Weights Matter.
What is the same inside every family
If there is one thing worth remembering, it is this: every family is built the same way, and that structure matters more than the vendor's name.
- A cheap, fast tier. For bulk work: classification, extraction, rephrasing.
- A middle tier. This is the workhorse — for most tasks it is the sensible choice.
- A top tier. For long multi-step work, hard code, serious derivation.
The price gap between the cheapest and the most expensive tier inside one family is often ten to twenty-five fold. Between vendors at the same tier it is far smaller. That is why picking the right tier saves more than switching vendor.
What the family name does not tell you
Whether it is good for your task. A family's top model can be excellent at code and mediocre at your language. The family name is a marketing category, not a capability guarantee.
Whether it is good in a smaller language. No vendor advertises this and few people measure it. You can test it yourself in five minutes — Which Models Handle Smaller Languages Well shows how.
Where your data runs. The same model may be available directly from the vendor, through a cloud provider and through an intermediary — three different contracts and three different data paths.
One number that rearranges the picture
The model router services publish how many tokens run through them on each model. That list does not look like the ordering you see in the news.
In September 2026, on one of the largest such intermediaries, the most tokens were consumed by a Chinese vendor's preview release, and four of the top five places went to open-weight, mostly Chinese models. The one Western entry near the top was not a flagship but the cheapest tier.
This does not mean those are the better models. It means that what developers run in bulk is the cheap thing — and that is not what the rankings measure. Leaderboards vs Real Use is about exactly that.
What to take away
Keep eight or ten families in your head, with one sentence on each: who makes it, what it is good at, whether the weights are open. Forget the version numbers — they will change anyway, and they are not what makes the decision; what you measure on your own task is.