Elorian's Glimpse AI zooms into images mid-reasoning — 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 and chat access, built around a "visual scratch space" that crops into image regions, pulls specific video frames, and runs code to measure or compare content mid-reasoning — rather than converting images to text tokens once and reasoning from memory.
- Glimpse was trained by a small team on a few hundred GPUs and is priced at $0.60 per million input tokens and $3 per million output tokens — cheaper than mid-tier models like Sonnet and Sol but more expensive than lightweight options like Haiku and Luna.
- Andrew Dai, Elorian's CEO and a former Google DeepMind researcher, said his motivation crystallized when a frontier model failed to identify who was winning a board game from a photo, since board games rely on visual iconography that text-token pipelines can't intuit.
- Elorian's internal evaluations show Glimpse outperforming GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning benchmarks — though the model is substantially smaller than the frontier systems it was tested against.
- Elorian came out of stealth six months ago with a $55 million seed round at a $300 million valuation, co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.
- Elorian is initially targeting robotics firms, scientific labs, media, and creative companies with heavy video workloads, and plans by next year to build what Dai called "the world's best visual reasoning model" — eventually moving from seeing to acting, including directly controlling robots.
Why it matters: Elorian bets the bottleneck in vision AI is visual grounding, not language reasoning, and is pricing Glimpse at $0.60/$3 per million tokens to undercut mid-tier competitors. If its benchmark claims against GPT-6 Astra hold, robotics and video-heavy companies get a purpose-built option that doesn't require frontier-model pricing — and the crop-and-zoom approach (already in OpenAI's o-series) gets validated as its own product category.
Ask SkimNews
