Elorian launches Glimpse AI model with visual scratch space — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Elorian released Glimpse as a research preview with API access, pitching a 'visual scratch space' that crops into regions, pulls video frames, and runs measurement code mid-reasoning rather than converting images to tokens once.
- Glimpse was trained by a small team on a few hundred GPUs, with Elorian pricing it at $0.60 per million input tokens and $3 per million output tokens — cheaper than mid-tier models like Sonnet and Sol, pricier than lightweight Haiku and Luna.
- Elorian's self-evaluated benchmarks claim Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks.
- CEO Andrew Dai, a former Google DeepMind researcher, told TechCrunch a frontier model couldn't identify the winner in a board-game photo because AI lacks the iconographic intuition humans take for granted on IKEA diagrams and game boards.
- Elorian raised its $55 million seed at a $300 million valuation before launch, co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital with Nvidia and Google chief scientist Jeff Dean participating.
- Elorian's crop-and-zoom mechanism resembles an approach OpenAI added to its o-series models in 2025, but the startup argues it built the entire model around that capability from scratch.
- Dai said Elorian's target customers sit on large video libraries — robotics firms, scientific labs, media, and creative companies — and that next year's goal is 'the world's best visual reasoning model,' eventually extending to robot control.
Why it matters: Elorian is entering a visual reasoning space where OpenAI's o-series already proved the crop-and-zoom approach works, but with a smaller model built cheaply from the ground up — targeting robotics and scientific labs that need video understanding at lower cost than frontier systems.
Ask SkimNews
