OpenRouter shows a free model that reads 1M tokens at one time

On November 3, 2026, OpenRouter added Crouton Alpha, a free model, with a tag that shows no maker. The model reads 1,048,576 tokens at one time, and it also reads images, audio and video. OpenRouter, a site for many AI models, did not make it and does not tell who did.

What it is

The model is for the small tasks that agents do again and again. Agents are software that does tasks for a person. It can write a maximum of 944,000 tokens. All prices are 0, but OpenRouter limits free use each minute and each day.

The numbers

First reports give a median speed of about 124 tokens each second. The maker and OpenRouter gave no benchmark scores. There is no model card, the document that gives test results.

What is not known

No person has found the maker. Posts on X name Mistral, xAI and a Chinese lab, but none is confirmed. It is not known if the model stays correct with very long inputs.

Insight: A long context window is not proof of quality

Crouton Alpha, LongCat 2.5 from Meituan and Mercury 3 from Inception all give a context window of about 1M tokens. None gave benchmark results for long inputs. This reads as a number that is free to show. Test results on long inputs set models apart. Test the model with your long input before you use it.