AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 16, 2026

Elorian, a Palo Alto AI startup that raised a $55 million seed round at a $300 million valuation before launching any product, released its first model on Friday, TechCrunch reported exclusively. The model is called Glimpse and arrives as a research preview: there is API access, while the chat interface is waitlist-only, so the company's public case for now rests on a results table and an argument about what current AI systems still cannot do.

That argument is that today's models, despite their fluency with language and code, do not genuinely see. Elorian's stated thesis is that most multimodal systems look at an image once, convert it into tokens, and then reason about it in text, roughly the way a person might describe a photo from memory rather than glance back at it. Glimpse instead returns to the source image repeatedly during reasoning. Mid-way through a chain of thought it can crop into a region, pull specific frames out of a video, or run code to measure and compare what it is looking at. Elorian calls this mechanism a "visual scratch space," and the analogy the company offers is someone pinching to zoom on a phone photo to double-check what they thought they saw.

The motivation came partly from CEO Andrew Dai, a former Google DeepMind researcher. He described photographing a finished board game after a game night last year and asking a frontier model who was winning; it could not answer. Board games, he said, each carry their own visual language, iconography and symbols, and while an adult can follow what is happening with minimal text, the models could not intuit what those symbols meant. He extended the point to Ikea diagrams, which people assemble with no text at all.

The technique is not entirely novel. OpenAI added crop-and-zoom during reasoning to its o-series models in 2025. Elorian's pitch is that it designed the entire model around the capability and did so cheaply: 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 is benchmarked against.

Pricing is set at $0.60 per million input tokens and $3 per million output tokens. That places Glimpse below mid-tier models such as Sonnet and Sol on cost, but above lighter options such as Haiku and Luna. On performance, 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. Those figures come from the company rather than independent testing, so they should be read as vendor-reported claims until outside benchmarks appear.

The company emerged from stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean. Its initial commercial targets are organizations that already sit on large volumes of video: robotics firms, scientific labs, media companies and creative studios. Longer term, Elorian wants to move from seeing to doing, building a model of a scene and using that representation to select actions, including direct control of robots. Dai said the company aims next year to build what he called the world's best visual reasoning model, one that leads on every visual reasoning test.

For developers, the practical shape of the offering matters as much as the benchmark table. A research preview with API access but a waitlisted chat interface is a common pattern for labs that want real usage data without committing to production-grade reliability, quotas or support. A model that interleaves cropping, frame sampling and code execution during inference also changes cost and latency profiles in ways a simple per-token price does not capture: each return to the source image is another step in the reasoning loop, and each code run adds execution overhead. Teams evaluating Glimpse should expect to measure effective cost per task rather than relying on the headline token rates.

This suggests Elorian is positioning against two different kinds of competitor at once. Against frontier multimodal systems, it argues that architectural focus on visual grounding beats raw scale. Against cheaper lightweight models, it argues that the scratch-space mechanism buys accuracy on genuinely visual tasks, such as reading a chart or captioning a short clip, that a text-centric pipeline handles poorly. A likely trade-off is that the approach shines on tasks with a clear visual ground truth and is less differentiated on tasks that are mostly language.

The deeper question is whether visual grounding generalizes into action. Robotics and lab automation are harder than chart reading because errors compound over time and the environment does not hold still. Elorian's roadmap bets that a model which can repeatedly inspect a scene is a better foundation for choosing what to do next than one that summarizes the scene once. That is an inference from the company's stated plans, not a demonstrated result.

Why it matters: teams working with video, charts or any visually dense content now have a cheaper, API-accessible option that claims state-of-the-art results on two specific visual tasks, though only Elorian's own numbers support that so far. The more consequential shift is architectural: if returning to the source image mid-reasoning becomes standard, evaluation and cost planning for multimodal features will need to account for iterative, tool-using inference rather than a single pass over an image. Product teams should treat the benchmarks as a prompt to test on their own data rather than a settled ranking.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

According to Elorian's evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks.