The Elorian Glimpse AI model launched on October 16 with a different approach to visual reasoning. Instead of analysing images once and relying mainly on text, it can repeatedly inspect visual evidence as it solves problems. The research preview includes an API and chat interface.
Elorian, a Palo Alto startup led by former Google DeepMind researcher Andrew Dai, developed the model after raising $55 million. The company previously secured a reported $300 million valuation before releasing its first product.
Many multimodal AI systems convert images into internal representations before reasoning about their contents. Glimpse can revisit those images, zoom into specific regions and examine individual video frames. It can also run code to measure or compare visual details, using what Elorian calls a visual scratch space.
How the Elorian Glimpse AI Model Handles Visual Reasoning
The system is designed to investigate visual information rather than depend entirely on its initial interpretation. For example, it can inspect specific video frames to track movement or revisit an image to verify an object’s position.
Dai said the limitations of existing models became apparent when he photographed a completed board game and asked an AI system to identify the winner. The model struggled to interpret the game’s symbols and layout. He argues that AI should eventually understand visual instructions as naturally as humans interpret furniture assembly diagrams.
Elorian detailed the approach in its official Glimpse announcement. The company says a small research team developed the model using a few hundred GPUs. It is substantially smaller than the frontier systems it compares against.
Glimpse costs $0.60 per million input tokens and $3 per million output tokens. That positions it below several premium AI models, although cheaper lightweight alternatives remain available.
According to Elorian’s evaluations, Glimpse surpassed GPT-6 Astra on scientific chart reasoning and short-form video captioning. It also performed competitively on several other visual tasks. However, Astra still led on certain perception and temporal reasoning benchmarks. These are company-reported results, not proof of superior performance across all workloads.
The underlying concept is not entirely new. OpenAI introduced crop-and-zoom capabilities during reasoning in 2025. Elorian’s distinction is its emphasis on incorporating repeated visual inspection throughout the model’s reasoning process.
Elorian Targets Robotics and Scientific Applications
Elorian’s initial customers are expected to include robotics companies, scientific laboratories, and businesses that process large video collections. Better visual understanding could help systems identify objects, interpret complex diagrams and follow actions across video sequences.
Other startups are pursuing related capabilities, including Perceptron, which is developing visual AI for industrial robots. Elorian eventually wants Glimpse to move beyond observation toward selecting actions and controlling physical machines.
The startup is backed by Striker Ventures, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. Dai plans to expandGlimpse’ss visual reasoning capabilities over the next year. For now, the research preview offers an early test of whether repeated visual inspection can make smaller AI models more capable.
