The models, what each costs, and how each was measured

Every model is asked through its maker's own API: OpenAI, Anthropic, Google and xAI directly, and the rest through OpenRouter. The rates are the provider's own; an answer costs its tokens and searches at these rates, times 2.5, and at least $0.001 (Pricing).

Offered models

"Cut off" is the share of this model's answers, across every poll in the last 30 days, that reached the length limit before answering. A model with few calls has a rough figure.

A model is offered only after it is measured

Before a model can go in a poll, we ask it a fixed set of multiple-choice and scale questions, several times each, and record how it answers: whether it gives an allowed answer within the 1,024-token limit for a short answer (an answer cut off there is asked once more, with twice the room, as in a poll), what each answer costs, and, with search on, how many searches it runs and what they cost. A model with an answer cut off even then is not offered. Its search is offered only at the depths it was measured at.

Through OpenRouter

Models reached through OpenRouter are served by one host each, chosen and pinned, and the host that served every answer is recorded with it, so an answer never silently comes from somewhere else.

Measured, not offered

• Mistral Large 4 (Mistral), measured Nov. 4, 2026: an answer was cut off even when asked again with twice the 1,024-token limit for a short answer

Showing these is how you can see where the bar sits.

When a model changes

Providers retire models and change their rates. A retired model keeps its history in every poll that asked it, and a new rate applies to answers from the day it takes effect.