The key facts

  • Elorian released its first model Glimpse as a research preview with API access; the chat interface is waitlist-only.
  • The model uses a 'visual scratch space': during reasoning it crops regions, extracts video frames, and runs code for measurements and comparisons.
  • Pricing: $0.60 per million input tokens and $3 per million output tokens — cheaper than Sonnet- and Sol-class models, more expensive than Haiku and Luna.

A model that 'looks' while reasoning

Elorian, a Palo Alto startup that raised a $55 million seed round at a $300 million valuation even before launching a product, released its first model on Friday — Glimpse, TechCrunch reports. The model is available in research preview status via API, while the chat interface works on a waitlist-only basis. The company's bet: modern models, for all their strength in language and code, don't truly 'see.' Most multimodal models process an image once, converting it into tokens, and then reason in text — like a person describing a photo from memory. Glimpse, however, returns to the original image: during reasoning it can crop a region, extract the needed frames from a video, or run code for measurements and comparisons. Elorian calls this mechanism a 'visual scratch space' — by analogy with a person pinching to zoom into a photo on their phone to check what they saw.

The founder's motivation

Elorian CEO Andrew Day, previously a researcher at Google DeepMind, told TechCrunch that one sign that AI cannot truly see was, for him, a game night: a year ago he photographed a board game after a session with friends and asked a frontier model who was winning — the model could not answer. According to him, board games each have 'their own visual language, their own iconography, symbols,' and an adult can easily understand what's happening with almost no text, while models cannot intuitively grasp the meaning of symbols. 'People can look at IKEA diagrams without text and assemble furniture. Why can't AI models do this?' Day said.

The approach isn't new, but cheapness is claimed

The idea isn't fully original: OpenAI added cropping and zooming during reasoning to its o-series models in 2025. However, Elorian claims it built the entire model around this mechanism and did so cheaply: according to the company, a small team trained Glimpse on a few hundred GPUs, and the model itself is substantially smaller than the frontier systems it is compared against in benchmarks.

Pricing and benchmark results

Glimpse costs $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. By Elorian's own assessments, Glimpse outperforms even GPT-6 Astra in reasoning over scientific charts and brief video summarization, showing competitive results on other visual reasoning tasks. It's worth noting that this data comes from the company itself.

Investors and plans

Elorian came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital; Nvidia and Google chief scientist Jeff Dean also participated. The initial target customers are companies with large volumes of video: robot manufacturers, scientific labs, media, and creative companies. In the long term, Elorian wants to move from 'seeing' to 'acting': building a scene model and using it to select actions, including directly controlling robots. According to Day, next year the company is aiming to create 'the world's best visual reasoning model,' leading in all benchmarks in this field.