Elorian’s first AI model wants to reason with images, not just about them

TechCrunch reports Palo Alto startup Elorian, which raised a $55 million seed at a $300 million valuation before launching a product, released Glimpse as a research preview with API and chat. CEO Andrew Dai, formerly of Google DeepMind, argues most multimodal models convert an image to tokens once and reason in text; Glimpse can mid-reasoning crop regions, pull video frames, or run code to measure what it sees.

Elorian says a small team trained Glimpse on a few hundred GPUs and that it’s substantially smaller than frontier systems it benchmarks against. Pricing: $0.60 per million input tokens and $3 per million output. Elorian’s evals claim Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning. Seed co-led by Striker, Menlo, and Altimeter, with participation from Nvidia and Google chief scientist Jeff Dean. Targets include robotics, scientific labs, media, and creative firms; longer term, Dai wants scene models that choose actions and control robots.

The board-game anecdote is the right critique: models that “see” once still miss visual languages humans parse instantly. Building the loop around re-looking is useful AI if the scratch space is cheap and reliable.

OpenAI added crop-and-zoom to o-series models in 2025, as TechCrunch notes — so the idea isn’t unique. Elorian’s bet is specializing early and staying smaller. We’ll wait for third-party chart and video benches before crowning it over Astra.

I’m watching API preview quality on scientific figures, and whether “seeing to doing” for robots stays roadmap talk or ships next year as Dai hopes.

Dai cites Ikea-style diagram understanding as a north star for visual reasoning without heavy text crutches.