We just shipped something a bit different: Notion Knowledge Board
Our goal is to measure how well different models 'get the job done' on real tasks.
We picked 15 models that passed our internal evals, split real production traffic across them, and used an ensemble of judges from Anthropic, OpenAI, and Google to score the results.
Quality came out close across all models: 94-98%, open source included. Cost varies a lot: $0.02 to $0.87 per task.
For most everyday work, cheaper models are now good enough. AI bills are climbing--and picking the right model is an easy way to spend less without losing much.