Elorian unveils Glimpse, a visual-reasoning model that can dynamically zoom and compute on images
The Palo Alto AI startup’s first model lets AI crop, pull video frames and run code during reasoning, targeting video-heavy industries.
Elorian, a Palo Alto-based artificial-intelligence startup, announced the release of its inaugural model, Glimpse, on Friday. Delivered as a research preview with both API access and a chat interface, the model is positioned as a response to a persistent gap in today’s multimodal systems: the inability to truly “see” and interact with visual content beyond a single pass tokenization.
Glimpse distinguishes itself by maintaining a “visual scratch space” that permits the model to revisit the original image while reasoning. Rather than converting an image once into static tokens and then describing it, Glimpse can dynamically crop regions, extract specific video frames, or even execute code to measure and compare visual elements mid-analysis. The capability mirrors a human’s instinct to pinch-zoom a photo to verify a detail, a functionality that Elorian’s CEO Andrew Dai, a former Google DeepMind researcher, highlighted after a board-game photo stumped a leading frontier model. Dai argued that humans effortlessly interpret visual symbols, from Ikea diagrams to game icons, while current AI often fails to grasp such context.
The startup notes that the approach is not entirely novel; OpenAI introduced a similar crop-and-zoom feature in its o-series models in 2025. However, Elorian claims to have built the entire architecture around this iterative visual reasoning, achieving it with a relatively modest compute budget. According to the company, a small team trained Glimpse on a few hundred GPUs, resulting in a model that is substantially smaller than the larger frontier systems it benchmarks against. Pricing reflects this efficiency: Glimpse costs $0.60 per million input tokens and $3 per million output tokens, undercutting mid-tier models such as Sonnet and Sol while sitting above lightweight offerings like Haiku and Luna.
In internal evaluations, Glimpse reportedly outperformed GPT-6 Astra on tasks involving scientific chart interpretation and short-form video captioning, and delivered competitive results across a broader suite of visual reasoning benchmarks. The startup emerged from stealth six months ago after securing a $55 million seed round at a $300 million valuation. The round was co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with strategic participation from Nvidia and Google chief scientist Jeff Dean.
Elorian’s immediate go-to market focus is on organizations that generate large volumes of video or visual data, including robotics firms, scientific laboratories, media producers, and creative studios. By offering a model that can both analyze and manipulate visual inputs, the company aims to streamline workflows that currently rely on manual annotation or separate vision and reasoning pipelines. Looking ahead, Dai envisions extending Glimpse from passive observation to active decision-making, ultimately enabling AI to construct scene models that inform real-world actions, such as controlling robots. The firm’s roadmap targets the development of “the world’s best visual reasoning model” by next year, aspiring to lead every standardized visual reasoning test.
If Glimpse delivers on its promises, it could reshape how founders and operators integrate AI into visual-intensive products, lowering barriers to sophisticated image and video understanding while opening new avenues for automation and insight generation. The coming months will reveal whether the model’s performance and cost advantages translate into broader adoption across the targeted sectors.