Most AI models that claim to “see” are actually just reading their own notes about an image. Elorian thinks that’s the core problem, and it has built its first model around fixing it. According to TechCrunch, the Palo Alto startup released Glimpse on Friday as a research preview, with API access available now and a waitlist-only chat interface. So yes, it’s still mostly a thesis and a results table. But the thesis is worth taking seriously.
The core critique Elorian is making is this: standard multimodal models look at an image once, convert it to tokens, and then reason entirely in text from that point forward. It’s the equivalent of glancing at a photo and then describing it from memory. Glimpse works differently. Mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to compare what it’s actually looking at. Elorian calls this a “visual scratch space.” Think of someone pinching to zoom on their phone to double-check something before they commit to an answer.
CEO Andrew Dai, a former Google DeepMind researcher, says the idea came partly from a board game moment. He photographed a finished game and asked a frontier model who was winning. It couldn’t tell him. Board games have their own visual logic, their own symbols, and current models can’t intuit what those symbols mean without heavy text support. His other example: IKEA instructions. Humans follow diagram-only assembly guides without much trouble. AI models still can’t.
The approach itself isn’t entirely new. OpenAI added crop-and-zoom reasoning to its o-series models in 2025. But Elorian’s argument is that it built the entire architecture around this capability, rather than bolting it on, and did it cheaply. The company says a small team trained Glimpse on a few hundred GPUs, and the model is substantially smaller than the frontier systems it’s being benchmarked against. Pricing reflects that: $0.60 per million input tokens and $3 per million output tokens. That puts it below Anthropic’s Sonnet and Inception’s Sol, but above lightweight options like Haiku and Luna.
On performance, Elorian claims Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. Those are specific, narrow categories, and the company’s own evaluations should be read with that caveat in mind. Independent benchmarking will matter a lot here.
Still, the market timing is reasonable. The startup raised a $55 million seed round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with Nvidia and Google chief scientist Jeff Dean also participating. Its initial targets are companies sitting on large volumes of video: robotics firms, scientific labs, and media companies. Longer term, Dai says the goal is to move from seeing to doing, including direct robot control. Next year, he wants Glimpse to lead every major visual reasoning benchmark. That’s an ambitious target. But the foundation they’re building on is at least architecturally distinct from what the big labs have shipped so far.
