Elorian, a Palo Alto-based artificial intelligence startup, has launched its first model, Glimpse, as a research release accessible through an API and a chat interface. The model focuses on reasoning based on the visual content itself, rather than converting an image or video into tokens once and then processing it through a fixed textual description.
Elorian says Glimpse can return to the visual source multiple times during the thinking process, crop a specific part of an image, extract particular frames from a video, or run code to measure and compare elements. The company calls this mechanism a “visual scratch space”, designed to allow the model to verify details that the initial analysis might miss.
From Description to Repeated Inspection
Andrew Dai, Elorian’s CEO and a former researcher at Google DeepMind, believes that current models’ inability to understand visual symbols and relationships appears even in everyday situations. He said he tested an advanced model with an image of a board game and asked it which player was ahead, but the model was unable to interpret the symbols and rules of the game visually. Dai compares this with humans’ ability to understand furniture assembly diagrams that contain little text.
The idea is not entirely new; OpenAI added cropping and zooming during inference to its models from the o series in 2025. However, Elorian says Glimpse was built around this approach from the outset while maintaining lower training costs. According to the company, the model was trained by a small team using a few hundred graphics processing units, and it is also much smaller than the leading systems against which it was compared.
Performance and Cost
According to Elorian’s evaluations, Glimpse outperforms GPT-6 Astra on scientific diagram reasoning and short-video description, and delivers competitive results on other visual reasoning tasks. These results come from the company’s own evaluations, and the source did not provide independent details about the test sets or comparison methodology, so they alone are insufficient to establish broad superiority over all models.
Glimpse costs approximately $0.60 per million input tokens and $3 per million output tokens. According to the comparison provided, this places it as less expensive than mid-range models such as Sonnet and Sol, but more expensive than lightweight models such as Haiku and Luna.
What Changes in Practice?
In the first phase, Elorian is targeting companies that handle large volumes of video, including robotics companies, scientific laboratories, and media and creative organizations. If independent testing confirms the company’s claims, the model’s practical value lies in reducing reliance on the initial description of an image and enabling more detailed examination of diagrams, symbols, and the sequence of events within a video.
Elorian raised a $55 million seed round at a valuation of $300 million before launching its product. Striker Venture Partners, Menlo Ventures, and Altimeter Capital participated in the round, along with Nvidia and Jeff Dean, Google’s chief scientist. Over the longer term, the company says it wants to move from vision to action by building a model of the scene and using it to select an action, including directly controlling robots. Dai also announced that the company aims to build “the world’s best visual reasoning model” next year, an ambitious goal that remains subject to results and tests whose details have not been fully disclosed.