Elorian, a startup whose chief executive, Andrew Dai, previously worked at Google Brain and DeepMind, released its first AI model, Glimpse, on Friday, Oct. 16. Most AI models that handle pictures first turn what they see into words and then work on the words, Elorian says, a chain it calls fragile and prone to mistakes. Glimpse is built to keep looking: while it works through a question it can zoom into part of an image, rewatch a short stretch of video more closely or follow one detail from frame to frame, and Elorian's chat site shows those close-ups to the user as Glimpse works.
Elorian's examples are everyday scenes. In one, the model follows a window on a train pulling into a station and, seeing it move a shorter distance between frames, concludes the train is slowing down. Another, the photo with this story, asks how many open pliers lie in a pile that also holds closed pliers and scissors. The answer Elorian checked is four, though the page does not show Glimpse's own attempt. The company says the skills it tested could help with jobs such as analyzing lab data and tagging social media videos.
Glimpse is an early preview, open now through Elorian's chat site and its API, a connection that lets developers plug a service into their own apps, and Elorian invites people to try it free. Developers pay $0.60 for every million tokens Glimpse reads and $3 for every million it writes; tokens are the small chunks of text, often parts of words, that AI models read and write. Elorian says that is 94% less than OpenAI's GPT-6 Astra and 85% less than Anthropic's Claude Opus 5.5.
Of the 11 tests in Elorian's own table that list both models, Glimpse beat GPT-6 Astra on four, including reading scientific charts, one of two counting tests and describing short videos people post online, and tied it on one. Astra scored higher on six, among them tests of understanding a scene, counting items in product photos and keeping track of exact timing in video, and Claude Opus 5.5 edged out Glimpse on the chart test. Elorian says it ran most models at their highest effort setting, though some video results used a lower one and some scene-understanding scores were taken from published papers.
Elorian says a small team built Glimpse on only a few hundred GPUs, the advanced chips AI models are trained and run on. Rather than start from scratch, it further trained an existing model whose trained files are published for anyone to download, and it has not said which one. The company says looking this closely takes extra time and extra tokens, the units developers pay for, and cutting that is its top priority next. It also wants to use what Glimpse sees to control robots, but says that work is still ahead.