Elorian Launches Glimpse: Visual Reasoning AI at $0.60/M Tokens

Executive Overview

Elorian, a Palo Alto-based startup, has unveiled Glimpse, its first AI model, as a research preview available via API and chat interface. Priced at $0.60 per million input tokens and $3 per million output tokens, Glimpse introduces a Visual Scratch Space that lets the model crop, zoom, and extract video frames during reasoning. The launch follows a $55 million seed round at a $300 million valuation, with backing from Nvidia and Google's Jeff Dean. Elorian claims Glimpse outperforms GPT-6 Astra in scientific diagram reasoning and short-video description.

Deep-Dive Features & Architecture

Glimpse is built on the thesis that current multimodal models "see" an image only once, converting it to tokens and then reasoning about it textually—like describing a picture from memory. Glimpse addresses this with a mechanism Elorian calls Visual Scratch Space, which allows the model during reasoning to return to the original source: crop a specific region, extract specific frames from a video, or run code to measure and compare what it sees. CEO Andrew Dai describes it as akin to pinching to zoom on a phone to verify what you thought you saw.

Dai cites a "game night" incident: he photographed a finished board game and asked a frontier model who won, but it couldn't answer because each game has its own visual language, icons, and symbols. He poses the question: "Humans can look at IKEA diagrams without text and assemble furniture, so why can't models?" Elorian claims a small team trained Glimpse on only a few hundred GPUs, and that the model is significantly smaller than the frontier systems it is compared to.

Benchmark & Competitive Performance

According to Elorian's internal evaluations, Glimpse outperforms GPT-6 Astra on two specific tasks: scientific diagram reasoning and short-video description, with competitive results on other visual reasoning tasks. On pricing, Glimpse sits in the mid-tier: cheaper than models like Sonnet and Sol, but more expensive than lightweight models like Haiku and Luna. Notably, OpenAI added crop-and-zoom during reasoning to its o-series models in 2025, but Elorian asserts it built the model entirely around this mechanism at low cost.

Industry Impact & Enterprise Adoption

At $0.60 per million input tokens, Glimpse becomes an attractive option for startups working on video analysis and visual content, especially in media, education, and robotics. The absence of an explicit Arabic support announcement means token efficiency for Arabic text will remain tied to the base model's performance in non-English languages, necessitating practical testing before production adoption. The most prominent advantage for developers is the ability to run code within the visual reasoning path, opening use cases such as reading engineering diagrams and documents.

Conclusion

Elorian's Glimpse represents a focused bet on visual reasoning as a distinct capability, backed by a lean training approach and aggressive pricing. While independent benchmarks are pending, the model's Visual Scratch Space and competitive claims against GPT-6 Astra position it as a notable entrant in the multimodal AI landscape. Enterprises with heavy video and diagram workloads should evaluate Glimpse's cost-performance trade-offs firsthand.

Media Source: TechCrunch AI | Fact Verification & Analysis: AI Tools Oasis