How we score AI models

What the score includes

The score is built from three parts: the model's capability (the main weight), reliability, and response speed. Capability drives most of the result; reliability and speed fine-tune it. Each part is put on a common scale, and the total is normalized to the current rating leader — so a 10 tracks the strongest model today, not an absolute ceiling.

How we assess capability

We assess capability from independent public benchmarks (general knowledge, coding, reasoning) and our own runs on real questions. Externally we show our own 0-10 composite, not the raw numbers of third-party benchmarks — it is more honest and avoids implying a precision no single test has.

Reliability and speed from our measurements

Reliability and speed come from our telemetry over the last 30 days: the share of successful answers and the median response time on live requests. A model without enough measurements gets the median of the rest — otherwise working flagships would be penalized by real numbers while unmeasured ones sat on a discount.

Price and context

We take price and context size from the open OpenRouter catalogue. The "cost per question ≈N ₸" on pages is an illustration for an average question profile; the real bill is computed from tokens actually spent, not from this estimate.

How often it updates

Prices, availability and capability scores are pulled automatically every hour. The "Maximum" and "Economy" line-ups are rebuilt from the fresh score, so model positions and page numbers change without manual edits.

What the score does not replace

The score compares models on general tasks — it does not guarantee the leader is best for your specific question. For important decisions (money, health, law), a single answer from any AI is worth double-checking. That is exactly what QueryWise does: it asks several strong models and shows where they agree and where they diverge.