Elorian launches Glimpse, an AI model that looks again

Elorian — the Palo Alto startup that raised a $55 million seed round at a $300 million valuation before shipping a single product — released its first AI model on Friday, TechCrunch has exclusively learned.

The model is called Glimpse. It arrives as a research preview, with API access open and the chat interface restricted to a waitlist. What buyers and researchers get today is a results table and a thesis. The thesis is blunt: today's AI models, for all their fluency with language and code, still can't really see.

That's a provocative claim in a market where frontier labs have spent years touting multimodal capabilities. But Elorian's founders think the industry has been solving the wrong problem.

What is a 'visual scratch space'?

Most multimodal models look at an image once, convert it to tokens, and then reason about it in text — like describing a photo from memory after a single glance. Glimpse keeps going back to the source. Mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to measure and compare what it's looking at. Elorian calls the mechanism a "visual scratch space."

The everyday analogy is simple: it works like someone pinching to zoom on a phone photo to double-check what they thought they saw. The model doesn't have to trust its first impression. It can verify.

This matters because the gap between "looking" and "seeing" shows up everywhere in practical deployments. A model that describes a warehouse shelf from a single pass can miss the one item that's out of place. A model that reasons over a chart from memory can misread the axis it never examined closely.

Why the CEO thinks AI still can't see

For former Google DeepMind researcher and Elorian CEO Andrew Dai, the moment of clarity came on game night. Last year he photographed a board game he had just finished with friends and asked a frontier model who was winning. It couldn't say.

Board games "each have their own visual language, each have their own iconography, symbols," Dai told TechCrunch. "To an adult, it's quite easy to understand what's going on with minimal text, but these models aren't able to intuit what the symbols mean."

He offered a second example that most readers will recognize instantly.

"Humans can look at IKEA diagrams, with no text, and assemble IKEA furniture, right? Why can't AI models?" he said.

The question cuts at a real limitation. Visual understanding that humans handle effortlessly — iconography, spatial layout, symbolic meaning — remains a stubborn weakness for systems that flatten an image into text-like tokens and then reason in words. If AI is to move from chatbots into robotics, scientific analysis, and physical work, that weakness becomes a bottleneck, not a curiosity.

Is the approach actually new?

Not entirely, and Elorian doesn't claim a monopoly on the idea. OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. The difference, according to Elorian, is architectural commitment: it built the whole model around the revisit-the-image mechanism rather than bolting it on.

There's also an efficiency argument. The company says a small team trained Glimpse on a few hundred GPUs, and that the model is substantially smaller than the frontier systems it benchmarked against. In an industry where frontier training runs consume tens of thousands of accelerators, a few hundred GPUs is a deliberate signal: capable visual reasoning may not require frontier-scale compute.

That framing puts Elorian in a growing camp of well-funded startups arguing that focused, efficiently trained models can beat general-purpose giants on specific task families — and that enterprises will buy specialized performance over brand-name generality.

How much does Glimpse cost, and how well does it perform?

Pricing lands in the middle of the market:

• $0.60 per million tokens in
• $3 per million tokens out

That makes Glimpse cheaper than mid-level models like Sonnet and Sol, but more expensive than lightweight models like Haiku and Luna.

On performance, Elorian's own evaluations claim Glimpse outperforms even GPT-6 Astra on two tasks: scientific chart reasoning and short-form video captioning. On other visual reasoning tasks, the company reports competitive results. These are vendor-run benchmarks, and independent validation will take time. But the claimed wins land precisely where the "visual scratch space" should help most — tasks that reward looking closely and repeatedly rather than describing from memory.

Who is backing Elorian?

Elorian came out of stealth six months ago. Its $55 million seed round was co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

That investor list reads as a bet on the thesis rather than the product. Nvidia's participation ties the startup to the compute supply it needs. Jeff Dean's involvement gives it a direct line to one of the most respected research minds in machine learning. Menlo and Altimeter bring enterprise distribution muscle. A $300 million pre-launch valuation, meanwhile, priced Elorian on the strength of its idea — Friday's release is the first hard evidence for whether that price made sense.

What comes after seeing?

The initial commercial targets are companies sitting on large volumes of video: robotics firms, scientific labs, media and creative companies. These are buyers whose workflows live or die on accurate visual interpretation, and who have the most to gain from a model that verifies what it sees instead of guessing.

The longer-term ambition is bigger. Elorian wants to move from seeing to doing — building a model of a scene and using it to choose an action, including directly controlling robots. That trajectory mirrors the industry's broader push from language models to embodied AI, where perception quality determines whether a robot grasps the right object or knocks over the shelf.

Dai has set a concrete milestone for it. Next year, he said, the company aims to build "the world's best visual reasoning model" — one that leads on every visual reasoning test. That's a tall order against frontier labs with vastly larger compute budgets. But if a few hundred GPUs can already put a small model ahead of GPT-6 Astra on chart reasoning, the benchmark race in visual AI may be more open than the language-model race ever was.

For now, Glimpse is a research preview, a waitlist, and a claim. The next year of benchmarks will tell us whether looking twice is the breakthrough the field has been missing — or a clever refinement on a path the giants already walk.