Elorian Launches Glimpse Visual AI, Claims GPT-6 Astra Wins

Summary

Elorian released its first model, Glimpse, on Friday, October 16, 2026, as a research preview with API access and a chat interface. Unlike multimodal models that convert an image into tokens once, Glimpse repeatedly revisits images and videos while reasoning, cropping regions, retrieving frames, and running code to measure or compare visual evidence in what Elorian calls a visual scratch space. CEO Andrew Dai, a former Google DeepMind researcher, cites frontier models’ inability to interpret board game symbols or text-free Ikea diagrams as evidence that AI still struggles to see.

OpenAI introduced crop and zoom reasoning in its o-series models in 2025, but Elorian says Glimpse was designed around the technique, trained by a small team on a few hundred GPUs, and is substantially smaller than benchmarked frontier systems. It costs $0.60 per million input tokens and $3 per million output tokens, less than Sonnet and Sol but more than Haiku and Luna. Elorian’s evaluations show Glimpse beating GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results elsewhere. The Palo Alto startup emerged from stealth six months ago after raising $55 million at a $300 million valuation in a seed round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with Nvidia and Google chief scientist Jeff Dean participating. Initial customers include video-heavy robotics, scientific, media, and creative companies. Dai aims to lead every visual reasoning test in 2027, then use scene models to select actions and directly control robots.