Elorian released its first model, Glimpse, on Friday as a research preview — and it matters because the Palo Alto startup built the whole system around re-examining images and video while it reasons, rather than glancing once and answering from memory.
October 16, 2026
What happened
Elorian came out of stealth about six months ago with a $55 million seed round that valued the company at $300 million, before it had a product to show. That round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with Nvidia and Google chief scientist Jeff Dean also taking part. On Friday the company finally shipped something: Glimpse, its first model, available through an API as a research preview. A chat interface exists too, but access there is waitlist-only for now, so the launch is mostly a set of evaluation results plus a clear statement of what the team thinks current models get wrong about vision.
How Glimpse approaches images differently
Most multimodal systems take in a picture, turn it into tokens, and then carry on reasoning in text alone. Elorian argues that this is where visual mistakes creep in: the model is effectively describing a photo from memory. Glimpse instead returns to the original material during reasoning. It can zoom into part of an image, pick out particular frames from a video, or run code to measure and compare details — a working area the company calls its visual scratch space. The inspiration the team points to is ordinary human behaviour: pinch, zoom, look again, then decide. OpenAI added a form of crop-and-zoom reasoning to its o-series in 2025, so the ingredients are not entirely new. Elorian’s claim is that it designed the entire model around that loop, and did so cheaply. A small team trained Glimpse on a few hundred GPUs, and the company says the resulting model is substantially smaller than the frontier systems it compares itself against.
Pricing, results and who it is for
API pricing is set at $0.60 per million input tokens and $3 per million output tokens. Elorian positions that below mid-tier models and above the cheapest lightweight options. On the company’s own evaluations — which are company-reported and not yet independently verified — Glimpse beat GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive scores on other visual tasks. The first customers Elorian is chasing hold large amounts of video: robotics firms, scientific laboratories, media companies and creative businesses. Longer term, the company wants to move from interpreting scenes to acting on them, including robot control, and chief executive Andrew Dai, a former Google DeepMind researcher, says the goal for next year is to lead every visual reasoning test the team tracks.
Why it matters
Vision is still the weak spot in a lot of everyday AI use. Models that sound confident about a chart, a board position or an instruction diagram often turn out to have misread the picture in the first place. A model that pauses to look again — cropping, measuring and checking before it answers — attacks that failure mode directly, and doing it with a smaller model trained on modest hardware would matter for cost as much as accuracy. The honest caveat is that Glimpse is a research preview backed mainly by the company’s own numbers. Until outside testers can run it against the same charts and videos, this is a promising thesis with a price tag attached, not a proven new leader. That is exactly the kind of claim worth testing — everything AI, tested.
Frequently asked questions
What is Elorian’s Glimpse model?
Glimpse is the first model from Palo Alto startup Elorian, released October 16, 2026 as a research preview. It is a multimodal model designed to re-examine images and video during reasoning instead of reasoning only from an initial text conversion.
How much does Glimpse cost?
Elorian lists API pricing at $0.60 per million input tokens and $3 per million output tokens. A chat interface is available on a waitlist basis.
How is Glimpse different from other multimodal models?
While it reasons, Glimpse can crop into image regions, select specific video frames and run code to measure what it sees — a process Elorian calls a visual scratch space. Most rival models convert an image to tokens once and then reason in text.
Has Glimpse’s performance been independently verified?
No. The results Elorian published at launch, including its claims against GPT-6 Astra on chart reasoning and video captioning, are the company’s own evaluations. Independent testing has not yet been reported.
Sources: TechCrunch