How Does Replit Get Frontier Performance at 40% of the Cost?

Amjad Masad says Replit gets frontier performance at 40% of the cost, pretty consistently, by giving customers the cheapest token that is still the best for the task. The benchmarks are published on the Replit website.

Replit trained that choice on three or four years of usage. The models pick the level a task needs, including model size and reasoning effort. In most cases, once a customer buys Replit, Replit wants them to trust it to select the model. Big enterprises and governments can still set restrictions. Zero-day retention contracts with the providers back that choice. With the freedom to choose, Replit stays at frontier performance and also significantly cheaper.