Elorian releases Glimpse, a model that re-inspects images while it reasons
Elorian, a Palo Alto AI startup that raised a $55 million seed round at a $300 million valuation before it had a product, released its first model on Friday, according to an exclusive TechCrunch report. The model, Glimpse, is launching as a research preview with API access and a chat interface.
The company's thesis is that today's AI models, for all their skill with language and code, still can't really see. Most multimodal models look at an image once, convert it to tokens, and then reason about it in text, which TechCrunch compares to describing a photo from memory. 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 or compare what it is looking at. Elorian calls this mechanism a "visual scratch space". The comparison offered is someone pinching to zoom on a phone photo to double-check what they thought they saw.
CEO Andrew Dai, a former Google DeepMind researcher, says one of the clearest signs came on game night. Last year he photographed a board game he had just finished with friends and asked a frontier model who was winning. It could not say. Board games, he told TechCrunch, "each have their own visual language, each have their own iconography, symbols", and while an adult can follow what is going on with minimal text, "these models aren't able to intuit what the symbols mean." He added that humans can assemble Ikea furniture from diagrams with no text, and asked why AI models cannot.
The approach is not completely new: OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian's pitch is that it built the whole model around the idea, and did it cheaply. The company says a small team trained Glimpse on a few hundred GPUs and that the model is substantially smaller than the frontier systems it is benchmarked against.
Glimpse costs $0.60 per million input tokens and $3 per million output tokens. That is cheaper than mid-level models like Sonnet and Sol, and more expensive than lightweight models like Haiku and Luna. According to Elorian's own evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks.
Elorian came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. The initial targets are companies sitting on lots of video: robotics firms, scientific labs, media and creative companies. Longer term, Elorian wants to move from seeing to doing, meaning building a model of a scene and using it to choose an action, including directly controlling robots. Next year, Dai said, the company aims to build "the world's best visual reasoning model", one that leads on every visual reasoning test.
Key facts
- Elorian, a Palo Alto startup with a $55 million seed round at a $300 million valuation, released its first model, Glimpse, as a research preview with API access and a chat interface.
- Glimpse can crop into image regions, pull frames from video and run code to measure or compare during reasoning, a mechanism Elorian calls a "visual scratch space".
- API pricing is $0.60 per million input tokens and $3 per million output tokens, between lightweight models like Haiku and Luna and mid-level ones like Sonnet and Sol.
- By Elorian's own evaluations, Glimpse beats GPT-6 Astra on scientific chart reasoning and short-form video captioning, and is competitive on other visual reasoning tasks.
- The company says a small team trained it on a few hundred GPUs; first customers are targeted among robotics, science, media and creative firms with lots of video.
Why it matters
Visual reasoning is the gap Elorian is going after: models that convert an image to tokens once and then think in text can miss details, as in Andrew Dai's board game example where a frontier model could not say who was winning. Glimpse is built so the model can return to the original image or video while it reasons. The idea is not new, since OpenAI added crop-and-zoom during reasoning to its o-series models in 2025, but Elorian says it designed the whole model around it and did so cheaply.
Who it affects
The stated first customers are companies sitting on lots of video: robotics firms, scientific labs, media and creative companies. Developers comparing multimodal APIs are also affected, since Glimpse is priced between lightweight and mid-level models. Longer term Elorian wants to move into choosing actions, including directly controlling robots.
How to use it
Glimpse is available as a research preview through an API and a chat interface. The API costs $0.60 per million tokens in and $3 per million tokens out, which is cheaper than Sonnet and Sol and more expensive than Haiku and Luna. The source does not say when the research preview becomes generally available or whether it is free.
How solid is it
The performance claims come from Elorian's own evaluations: Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results elsewhere. The benchmark names, scores and margins are not given, and no independent verification is mentioned. The training claim (a small team, a few hundred GPUs) and the claim that the model is substantially smaller than the frontier systems it is benchmarked against are also the company's. No parameter count is given. The funding and valuation figures are reported by TechCrunch, which learned of the launch exclusively.
Risks and caveats
Self-reported results on two tasks do not show the model leads across visual reasoning; on the rest, Elorian describes only competitive results. The company's stated goal for next year, to build the world's best visual reasoning model that leads on every visual reasoning test, is an aim, not a result. The robot-control ambition has no timeline in the source. The model is a research preview, so behaviour and terms may change.
“Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can’t AI models?”
— Andrew Dai, Elorian CEO, to TechCrunch