Elorian launches Glimpse, a visual-thinking AI model

Elorian released Glimpse on October 16, 2026, describing it as a visual-thinking model that reasons about the physical world from what it sees.[1][3] The company says Glimpse can work in visual space, revisiting and refining internal representations as its reasoning unfolds.[3] It is launching as a research preview with API access and a chat interface.[2] Elorian identifies robotics, scientific discovery and creative industries as areas where this kind of visual understanding could matter.[3] For teams considering a trial, the preview makes the model accessible through both a programmatic route and a chat UI.[2][3]

Elorian says Glimpse can crop into a region of an image while reasoning, pull specific frames from a video and run code to measure or compare what it is examining.[2] The company calls the inspection mechanism a “visual scratch space.”[2] Elorian also says the model was trained end-to-end to use multimodal tools that write code, transform images and watch videos.[3] It says Glimpse can follow moving objects, understand changing viewpoints and how parts fit together, and connect observations to what happens next.[3] These are company-described capabilities; the evidence does not establish how they perform in a particular workflow.

Glimpse is available through Elorian’s API and chat UI, with the company labeling the launch a research preview.[2][3] Elorian set the price at $0.60 per million input tokens and $3 per million output tokens.[3] The company says both rates are 94% lower than GPT-6 Astra’s.[3] Elorian also says a small team developed Glimpse using only a few hundred GPUs, and that it post-trained the model on an open-weights base model.[3] The company says improving token efficiency is a priority because its careful review takes time and tokens.[3]

Elorian says Glimpse exceeds Astra on four evaluations and matches it on embodied reasoning.[3] Its comparison table, however, records four wins, one tie and six losses against GPT-6 Astra across 12 rows, with one row lacking an Astra score.[1] Elorian also says Glimpse achieves frontier-level multimodal benchmark performance and sometimes exceeds GPT-6 Astra and Claude Opus 5.5, despite being substantially smaller and cheaper.[3] The supplied benchmark rows include results for CharXiv-R, UGC-VideoCap and open-vocabulary counting, but the evidence does not map the listed scores to model columns.[3] Teams should treat the comparison as Elorian’s reported results rather than a guarantee of performance on their own tasks.