Elorian has announced Glimpse, a new category of visual thinking model built to reason about the physical world from visual information. The system follows objects in motion, understands changing viewpoints, and connects observations to likely next events. The company describes Glimpse as an early step toward multimodal-native pretraining, generation, and direct robot control, with applications in robotics, scientific discovery, and creative industries.
Visual Thinking as a Core Capability
Unlike existing frontier large language models, Glimpse brings visual representations directly into the reasoning process. It can revisit and refine internal visual representations as thinking unfolds, making its understanding of objects and spatial relationships more grounded. As an agent, the model is also trained end-to-end to use multimodal tools that write code, transform images, and watch videos.
Coordinated Problem Solving
Glimpse can spawn and coordinate subagents that exchange text, visual artifacts, and internal representations to solve a shared problem. These capabilities allow the system to inspect a problem and revise its interpretation as it reasons. In the chat interface, users can observe the model's visual scratch space as it zooms into regions, watches short intervals at denser frame rates, or tracks features across time.
Frontier Benchmarks
On multimodal benchmarks, Glimpse reaches frontier-level performance at the level of GPT-6 Astra and Claude Opus 5.5, and it sometimes exceeds them. The model surpasses Astra on scientific chart reasoning, open-vocabulary counting, short-form video captioning, and manipulation-event understanding, while matching it on embodied reasoning. Elorian developed Glimpse with a small team using only a few hundred GPUs and post-trained it on an open weights base model.
Pricing and Efficiency
Glimpse costs US$0.60 per million input tokens and US$3.00 per million output tokens. These rates are 94% lower than GPT-6 Astra, 85% lower than Claude Opus 5.5, 70% lower than GPT-6.1 Sol, and 20% lower than Gemini 3.8 Flash introductory pricing. The pricing is intended to make visual reasoning more accessible for robotics, scientific discovery, and creative industries.
Applications in Robotics and Media
Elorian highlights applications including synthetic data generation for robotics training and System 2 planning for vision-language-action models. The model also supports dense captioning over short-form videos, precise video counting of objects such as puppies in a yard or equipment on a construction site, and simple visual reasoning puzzles. These capabilities target tasks that require following events, predicting outcomes from physics, and connecting actions to effects.
A Broader Push Toward Physical AI
Elorian states that building the full stack across systems, data, model architecture, and training from the ground up for multimodal intelligence can help unlock artificial general intelligence in the physical world. The company is working toward multimodal-native pretraining, models that both understand and generate, and direct action outputs for robots. Glimpse's visual understanding capabilities are an early step, while calibrated reconstruction and robot control remain work ahead.
Glimpse is available today through an API and an interactive chat interface. The platform shows the model's visual scratch space as it reasons, allowing users to observe zoomed regions, denser frame intervals, and tracked features over time. With frontier-level multimodal performance at lower cost, Glimpse offers a preview of Elorian's broader ambitions in physical intelligence.