Elorian releases Glimpse, an AI model that revisits images while reasoning

Elorian is promising AI that does more than describe what it sees. On October 16, it released Glimpse, its first model, with developer API access and a waitlisted chat interface. The research preview can revisit images while reasoning, but its claimed edge over larger rivals rests on Elorian’s own evaluations.

A second look, not just a longer answer

TechCrunch describes a common approach in multimodal AI—systems that handle images alongside text—as converting an image into tokens, then reasoning in text. Glimpse returns to the visual source during that process, using what Elorian calls a “visual scratch space.” Its inspection tools let it:

• Crop into a particular image region to examine it more closely.
• Pull specific frames from a video during the reasoning process.
• Run code to measure or compare what appears in the visual input.

CEO Andrew Dai told TechCrunch that a board-game photograph helped expose the problem he wanted to solve. Last year, he asked a frontier model who was winning a game he had finished with friends. It could not answer. He pointed to games’ distinct symbols and visual rules as information people can understand with little accompanying text.

Selected wins, not a universal lead

Returning to an image during reasoning is not itself new. OpenAI added crop-and-zoom capabilities to its o-series models in 2025. Elorian’s pitch is that it designed the whole model around this approach.

According to Elorian’s evaluations, Glimpse outperformed GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual-reasoning tasks. These are company-run comparisons, not independent findings.

Elorian also says a small team trained Glimpse using a few hundred GPUs, the processors used for AI training. It describes the model as substantially smaller than the frontier systems in its comparisons.

Glimpse costs $0.60 per million input tokens and $3 per million output tokens. TechCrunch places that below mid-level models such as Sonnet and Sol, but above lightweight options such as Haiku and Luna.

Video customers first, robot control later

The initial customers Elorian is pursuing are robotics firms, scientific labs, and media and creative companies—organizations working with large amounts of video.

Longer term, Elorian wants to model scenes and use that understanding to choose actions, including directly controlling robots. That is a development goal, not a capability established by this preview. Dai said the company aims in 2027 to lead every visual-reasoning test.