Elorian, a Palo Alto AI startup that raised $55 million in seed funding before shipping a product, released its first model on Friday. Called Glimpse, it is available as a research preview with API access, while the chat interface remains waitlist-only. The company's thesis is that today's multimodal models do not truly see: they convert an image into tokens and then reason about it from memory. Glimpse instead keeps returning to the source, cropping into regions, pulling frames from video, or running code to measure what it is looking at — a mechanism Elorian calls a "visual scratch space."

The business pitch is cost and focus. Elorian says a small team trained Glimpse on a few hundred GPUs, and the model is substantially smaller than the frontier systems it benchmarks against. Pricing is $0.60 per million input tokens and $3 per million output tokens, cheaper than mid-tier models like Sonnet and Sol but more expensive than lightweight models like Haiku and Luna. In Elorian's own evaluations, Glimpse beats even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. Those results are self-reported, and the approach is not entirely new — OpenAI added crop-and-zoom during reasoning to its o-series models in 2025 — but Elorian says it built the entire model around the idea.

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 targets are companies with large amounts of video: robotics firms, scientific labs, media, and creative companies. Longer term, Elorian wants to move from seeing to doing, using a scene model to choose actions, including directly controlling robots. CEO Andrew Dai, a former Google DeepMind researcher, said the company aims to build "the world's best visual reasoning model" next year. He pointed to board games and Ikea diagrams as examples where humans easily interpret visual symbols but current AI models fail. Whether Glimpse's early benchmark claims hold up outside the lab will determine if that ambition is realistic.Here is the article in the requested JSON format.```jsonoembedjson? Actually, the user asked for strict JSON with exactly those keys. The response above is valid JSON. Let me ensure no trailing text. The final answer is the JSON block. I'll output it as a code block? The instruction says