Elorian, an AI startup, released its first model, Glimpse, on October 16, offering it as a research preview with API access. The company states that Glimpse surpasses OpenAI's GPT-6 Astra in scientific chart reasoning and short-form video captioning, according to its internal evaluations. Elorian's pricing for Glimpse is $0.60 per million input tokens and $3.00 per million output tokens, which the company says is 94% lower than GPT-6 Astra's rates.
Glimpse distinguishes itself from other multimodal models by employing a "visual scratch space." This mechanism allows the model to revisit and refine its internal visual representations throughout the reasoning process. Most existing multimodal models convert an image into tokens once and then conduct all subsequent reasoning based on that initial textual representation. In contrast, Glimpse can crop into specific regions of an image, extract particular frames from a video, or execute code to re-examine visual content as it works to formulate an answer. Andrew Dai, Elorian's founder and a former Google DeepMind researcher, highlighted that this approach addresses limitations in current models, such as their inability to accurately interpret board game states or complex diagrams.
Elorian's evaluations show Glimpse achieving a 92.90 score on the CharXiv-R scientific chart reasoning benchmark, compared to GPT-6 Astra's 92.10. For short-form video captioning on the UGC-VideoCap benchmark, Glimpse scored 82.66 against Astra's 80.32. Glimpse also demonstrated higher performance on open-vocabulary counting (FSC-147) and manipulation-event understanding (WGO Bench). While Glimpse was competitive on other visual tasks, it trailed GPT-6 Astra on benchmarks such as HSS scene understanding and DREAM-1K dense captioning. Elorian notes that Glimpse is a substantially smaller model than the frontier systems it was benchmarked against, having been trained by a small team on a few hundred GPUs.
The company positions Glimpse as a "visual thinking" model, emphasizing its ability to understand how objects move, how parts fit together, and to connect observations to subsequent events. This capability is intended for applications in robotics, scientific discovery, and creative industries. Elorian also stated that Glimpse is trained end-to-end to use multimodal tools for coding, image transformation, and video analysis, and can coordinate subagents for problem-solving.
Elorian secured $55 million in seed funding at a $300 million valuation before the public release of Glimpse. The funding round included participation from NVIDIA, Striker Ventures, Menlo Ventures, and Altimeter. While Elorian's performance claims are based on its own internal evaluations, the company aims for Glimpse to lead all visual reasoning tests in the coming year. The chat interface for Glimpse is currently waitlist-only, though API access is available.