Great Work Labs

Prism 2: sharper image tags and image search

Jake Marsh ยท November 20, 2026

Prism 2 tags a photo with labels you choose and turns images and text queries into vectors for visual search, like Prism 1 does, only with far fewer mistakes. It takes the same options and returns the same fields, so switching is a change of model id. That brings the lineup to 35.

In our tests it tagged everyday photos with 0.883 mean average precision, up from 0.719 for Prism 1. It named the right breed for 94.7% of photos across 37 cat and dog breeds (Prism 1: 86.5%), and given a photo it ranked the right caption first 90% of the time (Prism 1: 75%).

Embeddings now have 768 numbers instead of 512, and they live in a different space from Prism 1's, so re-embed your images and text queries together when you switch. A call costs $0.00025 per image, and every 10 text queries add one more unit.

Prism 1 isn't going anywhere. Keep it when speed matters more than accuracy (it runs about 2.5 times faster), when your labels or queries aren't in English, or when you already have a library of Prism 1 embeddings. The bare name prism now follows the newest version, so pin prism-1 if you want to stay on it.

Try both in the playground with the same photo and labels.

RSS feed