Elorian ships first model Glimpse, built to reason with images rather than about them

The launch

Elorian, a Palo Alto AI startup that raised $55 million at a $300 million valuation before it had a product, has released its first model. According to TechCrunch, the model — called Glimpse — went live on Friday as a research preview with API access, while the chat interface is restricted to a waitlist.

The thesis behind Glimpse is that current AI systems, however fluent in language and code, effectively cannot see. Most multimodal models convert an image into tokens once and then reason about it entirely in text afterwards — analogous to describing a photograph purely from memory. Glimpse is designed to keep returning to the image itself while it thinks.

A visual scratch space

Mid-reasoning, Glimpse can zoom into a region of a picture, pull specific frames out of a video, or execute code that measures and compares what it is looking at. Elorian calls this mechanism a "visual scratch space"; TechCrunch compares the behaviour to a person zooming into a photo on their phone to make sure they really saw what they think they saw.

CEO Andrew Dai, a former Google DeepMind researcher, told TechCrunch that the shortcoming became obvious at a games night: he photographed a board game he had just finished playing and asked a frontier model who was winning, and it could not say. Board games carry their own symbols and visual language that adult humans parse with little text, he argued, noting that people manage to assemble furniture from text-free Ikea diagrams.

The core idea is not unprecedented. As TechCrunch points out, OpenAI added crop-and-zoom behaviour during reasoning to its o-series models in 2025. Elorian's claim is that it designed the whole model around this loop, and did so cheaply: a small team, a few hundred GPUs, and a model the company says is substantially smaller than the frontier systems it benchmarks against.

Pricing and benchmarks

Glimpse costs $0.60 per million input tokens and $3 per million output tokens, per the pricing TechCrunch reports. That places it below mid-tier models such as Sonnet and Sol, but above lightweight options like Haiku and Luna.

On capability, Elorian's own evaluations put Glimpse ahead of GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. These are vendor numbers rather than independent findings — a caveat worth holding onto given the model is arriving as a preview with no widely available chat interface.

Who it targets, and what comes next

Elorian's initial customers are organisations sitting on large volumes of video: robotics firms, scientific labs, media businesses and creative studios. Longer term, the company wants to move from seeing to doing — building a model of a scene and using it to choose actions, including directly controlling robots. Dai said that next year Elorian aims to build what he calls the world's best visual reasoning model, one that leads on every visual reasoning test.

Elorian emerged from stealth six months ago with a seed round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

Why it matters

The launch tests a specific hypothesis: that the bottleneck in multimodal AI is visual grounding during reasoning, not simply model scale. If Elorian's benchmark claims survive outside scrutiny, a small team on a few hundred GPUs will have matched frontier-lab behaviour on visual tasks — an uncomfortable data point for labs spending orders of magnitude more on training runs. For industries drowning in video that language-first models handle poorly, a cheaper API built for repeated, code-assisted inspection of imagery is a usable option today, with robotics as a more ambitious second act. The obvious caveats are the standard ones for a research preview: self-reported evaluations, a waitlist-only interface, and no independent verification yet.