This site names five model families — DeepSeek, Qwen, GLM, MiniMax and GPT-OSS — and never a version number. That is deliberate, and this page explains why and what it means for you.
Why families, not versions
A version number tells you almost nothing useful and goes stale almost immediately. Providers replace the model behind a name, change its default settings, retire an old one at short notice and occasionally roll a change back. If we printed a version, it would be wrong within weeks and you would have no way of knowing. A family name is stable, honest about what it is, and enough to make sense of a comparison: you can see that one family is consistently more cautious than another without needing to know which build produced the sentence.
It also keeps the comparison about behavior you can see on the page rather than about specifications you cannot check. Everything we publish about the families comes from runs on this site.
Providers change without notice
The models are hosted by third parties. They may be updated, slowed down, rate-limited or withdrawn at any time, and we find out the same way you do — a family suddenly answers differently, or stops answering. When a family cannot answer, its segment says so plainly and you can retry that one on its own. We do not silently swap in a different family to fill the gap.
How a family is switched on or off
Each family occupies a fixed slot with its own settings: whether it is on, whether it is one of the defaults, how long an answer it may write, and how expensive it is relative to the others. Turning a family off removes it from the site completely — no chip, no segment, no mention. Three families answer by default because that is a good balance between a useful comparison and the cost of every question; the other two are one tap away and stay switched on for your next visit.
If a family is paused because the month’s budget is nearly spent, its segment says “paused this month” rather than pretending to try.
Thinking before answering
Some families can work through a problem privately before they start writing. Where that mode exists, we usually keep it off: it makes answers slower and more expensive without making them noticeably better for everyday questions, and a long silent pause is a poor experience when four other families are already writing. When a family does think first, its segment says “reasoning” instead of “thinking”, and only the final answer is shown — the private working is not part of the answer and is not stored.
What the site stores about you
A first-party cookie holds a random identifier so the daily counter works, and that is the whole of it. No account, no email address, no profile, no cross-site tracking from us. Questions and answers are cached by wording for seven days so repeats are instant; that cache is not linked to the cookie. Aggregate figures — how many questions were answered, how long each family took, how many people marked an answer as best — are kept without anything that identifies a person, and the monthly Agreement Report publishes them.