How Good Is GPT-6 Astra at Vision? We Tested It on 4,000 Images

​Large frontier models, like GPT-6 Astra, are becoming better and better at vision. But how much have they improved, exactly? And in which domains do they excel? Let's take a look at the benchmarks.

​Matvei Popov, Machine Learning Engineer at Roboflow, will introduce the RF100-VL benchmark, which is 100 datasets from real projects like construction sites, X-ray scans, shelves of retail product, and more. He will share findings from running GPT-6 Astra against a selection of those datasets, evaluate how it performed in zero-shot and few-shot settings, and explain where it outperformed other models. You will see which datasets Astra is strongest on, which ones still challenge it, and what that means when you are picking a model.

​We'll also look at how frontier models can be used in production vision systems. That includes examples we've seen using GPT-6 Astra to assist with developing smaller, specialized vision models, and how it might be chained with other models in Workflows.