Best AI models for summarisation

What to weigh for summarisation

Summarisation reads a lot and writes a little, so it is priced by the input rate and judged on whether the summary is faithful. Models above $5 per million tokens blended are left out of this list: for most summaries they are not better in a way a reader notices, and they cost five to twenty times more to read the same input.

The intelligence index remains the right ranking inside the band because faithfulness tracks it: the failure mode of a cheap summary is an invented detail, and the better-scoring small models invent less. Context window matters only when the input is long; the long-documents list covers that case.

The router's contribution is to notice which summaries are hard. A ten-page technical report with figures scores high on the complexity signal and goes to the strongest model in the band; a support ticket does not, and goes to the cheapest. Outcomes - a summary rated poor - move the next similar request up. Summaries are also the task where a request-level cache pays: the same article summarised for three readers is one model call, and the middleware that does it is one line in the router's configuration.