Artificial intelligence has made significant progress in understanding language, generating content, and writing code. However, interpreting visual information accurately remains a challenge for many AI systems. Palo Alto-based startup Elorian is looking to address this limitation with its first AI model, Glimpse, which is designed to analyze images and videos more actively rather than relying on a single initial interpretation.
The company introduced Glimpse as a research preview on October 16, 2026, providing API access while its chat interface remains available through a waitlist. Its central idea is to help AI systems examine visual details repeatedly during the reasoning process, potentially improving their ability to understand complex scenes.
Moving Beyond Basic Image Recognition
Many multimodal AI systems process an image by converting its visual information into tokens before generating a response. Although this approach works well for numerous everyday tasks, it can struggle when an answer depends on small details, unfamiliar symbols, precise measurements, or relationships between objects.
Glimpse takes a different approach. During its reasoning process, the model can revisit the original visual input, zoom into selected areas, extract particular video frames, and use code to measure or compare elements.
Elorian describes this capability as a visual scratch space. Similar to how a person zooms into a photograph to inspect a small detail, the model can gather additional visual evidence before reaching a conclusion.
This approach could prove useful for tasks where an initial overview is insufficient and accurate interpretation requires closer examination.
Why Visual Reasoning Matters
Elorian CEO Andrew Dai, a former researcher at Google DeepMind, has highlighted the limitations of current AI models through practical examples.
One example involved a board game. After photographing a completed game, Dai asked an advanced AI model to identify which player was winning. The model struggled to interpret the board’s symbols and visual conventions.
For humans, understanding a board game, following an illustrated assembly guide, or interpreting a scientific diagram often requires little written explanation. People combine visual patterns, contextual clues, and prior knowledge to understand what they see.
AI systems can struggle with these tasks because recognizing individual objects is not the same as understanding how visual elements work together.
Glimpse aims to narrow that gap by allowing the model to inspect evidence more carefully while solving a problem.
Performance, Development, and Pricing
Elorian says it developed Glimpse with a relatively small team and a few hundred GPUs. The company also positions the model as substantially smaller than the leading AI systems used for comparison, suggesting that visual reasoning capabilities may not always require the largest available models.
According to Elorian’s own evaluations, Glimpse outperformed GPT-6 Astra on scientific chart reasoning and short-form video captioning. The company also reported competitive performance across other visual reasoning tasks. These are vendor-reported results and should be considered alongside independent testing.
The model’s published pricing is $0.60 per million input tokens and $3 per million output tokens. This places it between some lightweight AI models and more expensive alternatives, depending on the workload and provider.
Elorian emerged from stealth approximately six months before the launch, following a $55 million seed funding round at a reported $300 million valuation. Its investors included Striker Venture Partners, Menlo Ventures, Altimeter Capital, and Nvidia, alongside participation from Google chief scientist Jeff Dean.
Industries That Could Benefit From Glimpse
Elorian is initially targeting organizations that work with large volumes of visual data.
Potential applications include:
- Robotics: Interpreting surroundings and identifying objects before taking action.
- Scientific research: Examining charts, experimental images, and complex visual measurements.
- Media and entertainment: Analyzing video footage and identifying important scenes.
- Creative industries: Understanding visual compositions and supporting image-based workflows.
The ability to revisit visual evidence could be particularly valuable when a small detail changes the interpretation of an entire image or video.
From Understanding Images to Taking Action
Elorian’s long-term ambition extends beyond visual analysis. The company wants to develop AI systems that can build an understanding of a scene and use that information to determine appropriate actions, including controlling robots directly.
This direction could eventually connect visual reasoning with physical-world automation. However, reliable interpretation, decision-making, and safety will remain important challenges as these systems become more capable.
The key takeaway: Glimpse represents an emerging approach to multimodal AI that emphasizes active visual investigation instead of one-time image interpretation. If its capabilities hold up under independent evaluation, it could help businesses extract more meaningful insights from images and videos while opening new possibilities for research, robotics, and intelligent automation.
