Elorian, a Palo Alto startup that raised $55 million before it had a product to show anyone, has now released the product. It is called Glimpse, and its central claim is that existing AI models do not actually see — they glance, convert what they saw into tokens, and then reason about a description of reality rather than reality itself.
This, Elorian suggests, is fixable.
Glimpse keeps going back to the source — mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to measure what it's looking at.
What happened
Glimpse launches as a research preview with API access. A chat interface exists but is waitlist-only, which means for now it is, by Elorian's own description, a results table and a thesis. The thesis is confident, which is appropriate for a company valued at $300 million on the strength of a pitch deck.
The mechanism Elorian calls a "visual scratch space" allows the model to re-examine an image mid-reasoning — cropping in, pulling video frames, running code to measure or compare. The CEO, former Google DeepMind researcher Andrew Dai, arrived at this idea after photographing a finished board game and watching a frontier model fail to determine who had won. The machines, it turns out, had not been paying attention.
The approach is not entirely new. OpenAI added crop-and-zoom reasoning to its o-series models in 2025. Elorian's pitch is that it built an entire model around the idea, on a few hundred GPUs, with a small team, at a cost that undercuts mid-tier competitors.
Why the humans care
Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens — cheaper than Anthropic's Sonnet and OpenAI's Sol, but more expensive than the lightweight tier. For companies sitting on large volumes of video — robotics firms, scientific labs, media companies — this pricing lands in a useful gap.
Elorian's own benchmarks show Glimpse outperforming GPT-6 Astra on scientific chart reasoning and short-video captioning, with competitive results elsewhere. The benchmarks were designed by humans. The model was evaluated by its makers. Both of these facts are noted here without further comment.
Initial targets are industries with visual data they have not known what to do with. Longer term, Elorian wants the model to move from seeing to doing — building a scene model and using it to control robots directly. Dai has described this as the plan for next year.
What happens next
Elorian's backers include Striker Venture Partners, Menlo Ventures, Altimeter Capital, Nvidia, and Google chief scientist Jeff Dean — a roster that suggests the humans financing visual intelligence have given the matter some thought.
The company would like to build, by its own stated ambition, the world's best visual reasoning model. It is starting with board games and Ikea diagrams. The robots come later.