Not reasoning from a single look
Most multimodal models turn an image into tokens and then do the rest of their reasoning largely in text. Glimpse, which Elorian released on October 16, takes a different approach: when it is unsure mid-reasoning, it goes back to the source image, zooms into a region, or rewatches a stretch of video at about 14 frames per second instead of about 8. It can also write code to measure or transform images and spawn subagents to split up the work. Elorian calls this working area a "visual scratch space" and says the tool use was trained end to end.
Self-reported results, wins and losses
All of Elorian's comparisons are its own, run at maximum thinking effort with each model's native agent harness. Against GPT-6 Astra, Glimpse leads on four: WGO Bench (42.70 vs 39.42), CharXiv-R (92.90 vs 92.10), UGC-VideoCap (82.66 vs 80.32) and FSC-147 counting (83.36 vs 82.02). It ties on ERQA and trails on six, with the widest gaps on Video-MME v2 with subtitles (56.68 vs 64.06) and HSS (46.36 vs 53.6). On CharXiv-R, Claude Opus 5.5's 93.40 is also higher than Glimpse.
Price is the main pitch
Pricing is $0.60 per million input tokens and $3 per million output tokens. In Elorian's robotics example, using Glimpse as the planner on a LIBERO task cost $0.035 per episode versus $0.249 for Astra, and both completed the task. Glimpse was post-trained on an open-weight base model by a team using only a few hundred GPUs.
The limits of a preview
The company says plainly that looking this closely takes time and tokens, and that token efficiency is its next focus; generating robot actions directly is still future work. According to TechCrunch, Elorian's CEO is former Google DeepMind researcher Andrew Dai, the company raised a $55 million seed round, and its target customers are robotics firms, scientific labs, and media and creative companies with heavy video workloads.