TL;DR
Elorian's first model, Glimpse, is a research preview focused on visual reasoning. The company reports competitive benchmark results and lists token prices of $0.60 per million input tokens and $3 per million output tokens. Its claimed 94% discount refers to token prices, not guaranteed savings per completed task.
The focus: visual reasoning
Elorian focuses on visual reasoning. Its stated goal is to improve how models interpret spatial relationships and physical constraints, areas the company argues remain difficult for systems centered on text.
Glimpse can inspect images and video during a reasoning task, using tools and a visual scratch space to check details. Elorian also describes subagents that share visual artifacts while working together.
A game-night motivation, with precedents
TechCrunch reports that a board-game scoring failure helped motivate CEO Andrew Dai's focus on visual reasoning.
Revisiting visual evidence during reasoning has precedents. OpenAI introduced native image manipulation with o3 and o4-mini in April 2025. Its release described cropping, zooming, and rotating images during reasoning.
What the numbers say — and what they don't
On Elorian's own results table, Glimpse beats GPT-6 Astra on four evaluations: scientific chart reasoning (CharXiv-R 92.90 vs 92.10), open-vocabulary counting (FSC-147 83.36 vs 82.02), short-form video captioning (UGC-VideoCap 82.66 vs 80.32), and manipulation-event understanding (WGO Bench 42.70 vs 39.42). It ties Astra on embodied spatial reasoning (ERQA 83.75 each).
Astra leads on several other evaluations in the same table. These are publisher-reported results using each model's native harness. The methodology notes exceptions for video reasoning settings and incorporates paper-reported HSS results. The performance figures discussed here come from Elorian's announcement.
Token prices and practical costs
Elorian lists per-million-token rates of $0.60 input and $3 output. Its 94% comparison with GPT-6 Astra concerns token prices, not guaranteed savings per completed task. A useful cost comparison also needs the number of tokens used and the quality of the completed work.
Glimpse is available as a research preview through an API and chat interface. Elorian identifies robotics, scientific work, and media as potential applications; those uses still need evaluation against the requirements of each workflow.