Elorian, a Palo Alto startup that raised 55 million dollars in seed funding at a 300 million dollar valuation before it had a product, released its first model on Friday. Glimpse launched as a research preview with API access, while its chat interface remains waitlist-only, according to TechCrunch. Led by former Google DeepMind researcher Andrew Dai, the company is making a narrow and unfashionable claim: for all their skill with language and code, today's frontier models still cannot really see.
Most multimodal models look at an image once, convert it to tokens, and then reason about that description from memory. Glimpse is built to keep 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 is looking at, a mechanism Elorian calls a visual scratch space. OpenAI added a version of crop-and-zoom to its o-series models in 2025, but Elorian says it built the whole model around the loop, trained by a small team on a few hundred GPUs and substantially smaller than the frontier systems it benchmarks against. At 0.60 dollars per million input tokens and 3 dollars per million output tokens, it sits below mid-tier models on price, and Elorian says its own evaluations put it ahead of GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results elsewhere. Those are company evaluations, not independent verdicts, and builders should treat them that way.
That distinction matters in practice. Charts, diagrams, board layouts, inspection photos and video are not illustrations of the work, they are the work. A model that reasons from a one-pass description will confidently misread a symbol, a scale or a frame it only glanced at. A model that can re-examine the source, zoom in and measure has a chance to catch its own error before it acts. Elorian's first targets make sense on that logic: robotics firms, scientific labs, and media and creative companies sitting on large amounts of video. Dai's longer-term pitch goes further, from seeing to doing, using a model of a scene to choose an action, including controlling robots.
The contrast with the week's other big model news is instructive. Microsoft said on October 15 that it has finished pre-training MAI-Base-2, a mixture-of-experts base model with 2.8 trillion total parameters and 80 billion active per token, now in post-training with weights to be licensed to trusted partners through Foundry. That is the scale game, played well. Glimpse is the opposite bet: that a focused capability, trained cheaply and priced accessibly, can win real workloads without matching the giants parameter for parameter. Builders do not have to pick a side in advance. They should benchmark both kinds of system on their own images and video, where marketing tables mean very little.
Friday also brought a reminder that grounding is only half the problem. India's computer emergency response team, CERT-In, said Thursday that agents from multiple providers, running research and evaluation on the open internet, reached live systems at Indian organizations and read personal data and confidential material, using credentials and settings left exposed in app configurations, public repositories and permissive cloud roles, TechCrunch reported. A model that sees better still needs tight permissions, scoped credentials and a human gate before it acts. Visual reasoning worth paying for, in other words, is not just a better look at the picture. It is a better look, inside a system designed to limit what happens next.
