Parameters, and big-model smell

If you’ve ever wondered why one model feels smarter than another when the benchmarks say they’re neck and neck, this is the concept you’re missing. It’s also the theory behind our Overall podium, so it’s worth ten minutes.

The intuition

A model’s parameters are the numbers it learned during training - the dials, billions of them, that got tuned by reading a large fraction of everything humans have ever written. More parameters means more capacity to store patterns: not facts in a database sense, but compressed structure about how the world works. How legal contracts differ from marketing copy. Why a pricing decision touches a UX decision. What a sensible person would do next.

You don’t need the maths. The useful mental model is this: a small model knows the road it was trained to drive; a big model has seen the whole map. Both get you to the shops. Only one copes when the road is closed.

Total vs active - the bit everyone gets wrong

Modern frontier models are mostly ‘mixture of experts’ designs: a huge total parameter count, of which only a fraction fires on any given token. That firing fraction is the active parameter count, and it’s the one that matters for the thing we call intelligence-in-the-moment. Two models can have similar totals and wildly different active counts - and the one activating more of itself per token is doing more thinking per word, and costing more per word too. That’s the trade, and it’s why the biggest models are expensive in a way that never quite goes away.

Labs have mostly stopped publishing these numbers, so everything downstream is inference. But the inference is decent: pricing, speed, and behaviour under ambiguity all leak information about size.

The house theory

Our Overall category ranks what we call big-model smell: judgement, taste, cross-domain synthesis - the ability to make the call a benchmark can’t score. The house theory is that big-model smell tracks active parameter count almost directly. Coding benchmarks reward narrow, RL-sharpened competence, and mid-size models increasingly saturate them. Judgement doesn’t saturate. When a task has undecided trade-offs - business versus UX versus maintenance - the model with more of the world firing per token simply has more to reason with.

It’s a theory, not a law, and it’s falsifiable: a small model topping blind human preference across open-ended work would break it. Nothing has yet.

What this means for your model choice

• Well-defined task, clear spec: parameters are wasted money. Use the cheap tier and route hard.
• Ambiguous brief, judgement calls, anything customer-facing where tone matters: pay for the big model. The failure mode you’re avoiding is not ‘wrong answer’ - it’s ‘confidently plausible answer that a human would never have chosen’.
• Ignore total-parameter marketing. Behaviour under ambiguity is the only size signal you can actually observe.

For how size interacts with what you pay per request, read The tokenizer tax next - the two concepts together explain most of the board.