How Elorian launched on X

Glimpse is the first model from Palo Alto lab Elorian, released as a research preview on 16 October 2026. It reasons about images and video by following objects through motion and changing viewpoints, judging how parts fit together, and linking an observation to what happens next. During reasoning it keeps a visual scratch space, feeding crops, selected frames, and new views back into later steps, and it can write code, transform images, and watch video, including by coordinating subagents. Elorian is aiming it at robotics, laboratory analysis, and media work such as reading scientific charts, captioning short video, and identifying completed actions in robot footage. On the company's own evaluations, Glimpse scores above GPT-6 Astra on CharXiv-R scientific chart reasoning (92.90 to 92.10), FSC-147 open-vocabulary counting (83.36 to 82.02), UGC-VideoCap short-form video captioning (82.66 to 80.32), and WGO Bench manipulation events (42.70 to 39.42), and ties Astra on ERQA embodied spatial reasoning at 83.75. A small team post-trained the model on an open-weights base with a few hundred GPUs. Input tokens cost $0.60 per million and output tokens $3.00 per million, which Elorian says is 94 percent below Astra. Glimpse is available through the company's API and chat UI. Elorian was co-founded by CEO Andrew Dai, who spent 12 years at Google Brain and DeepMind, including as Gemini data lead and a co-lead of PaLM 2 pretraining, Yinfei Yang, who led Apple's MM1 models and co-created Google's ALIGN, and Seth Neel, a former Harvard professor. The company left stealth in April 2026 with a $55 million seed round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Jeff Dean. Glimpse is the first product from that round.