The Pulse
Elorian Launches Glimpse to Reason Through Images and Video
Elorian has released Glimpse as a research preview, describing a model that can revisit visual evidence while reasoning about images and video. The company reports results across multimodal evaluations and says direct action outputs for rob
AI.info Team · Written with AI ·
“Humans can look at IKEA diagrams, with no text, and assemble IKEA furniture, right? Why can’t AI models?”
Andrew Dai, chief executive of Elorian, speaking to TechCrunch
Elorian released Glimpse on October 16 as a research preview, pitching it as a model that can inspect visual evidence again while reasoning instead of relying only on an initial image-to-text conversion. The company says the model can follow objects as they move, compare changing viewpoints and infer what may happen next. API access and an interactive chat interface are listed as available; TechCrunch reported that the chat interface was waitlist-only at launch.
Glimpse returns to the visual evidence
Many multimodal systems translate an image into tokens, then work through a question in text. Elorian says Glimpse can revisit and refine internal visual representations during reasoning. Its “visual scratch space” can include zooming into an image region, examining selected video frames at a denser rate or running code to measure and compare what appears.
The model also uses multimodal tools to write code, transform images and inspect video, according to Elorian’s launch post. The company says Glimpse can coordinate subagents that exchange text, visual artifacts and internal representations. Those features are intended to let it check an interpretation against the source material before answering.
Elorian reports gains on four evaluations
Elorian says Glimpse exceeds GPT-6 Astra on four evaluations: scientific chart reasoning, open-vocabulary counting, short-form video captioning and understanding manipulation events. On its chart-reasoning test, Glimpse scored 92.90 against Astra’s 92.10; on short-form video captioning, the scores were 82.66 and 80.32. The company reports that the models matched on embodied spatial reasoning, with both scoring 83.75.
Those are Elorian’s own results, not an independent evaluation. Its comparison table says results generally use each model’s highest thinking effort and a native agent setup, with different tool configurations for different systems. For video evaluations, Elorian says GPT-6.1 Sol and Opus 5.5 used high thinking effort, except Sol’s Video-MME v2 result, which used max; max-effort results are unavailable for the other video evaluations. Results vary across tests: Astra scored higher than Glimpse on the company’s fine-grained visual reasoning, dense video captioning and temporal reasoning entries.
Elorian lists prices of $0.60 per million input tokens and $3 per million output tokens. Its table compares those rates with $10 and $50 for GPT-6 Astra, respectively. The company says a small team trained Glimpse using a few hundred GPUs, then post-trained it on an open-weights base model.
Robot control remains a stated next step
Elorian points to robotics, scientific research and media work as potential uses, including labeling robot-training video and searching large video collections. Its launch examples include using Glimpse as a planner for a robot policy alongside GPT-6 Astra. In that example, both planners complete the task; Elorian reports a cost of 3.5 cents for Glimpse’s planning episode and 24.9 cents for Astra’s.
The company draws a boundary between interpreting scenes and directly acting in them. Its post says it is developing direct action outputs for robots; calibrated reconstruction and robot control remain work ahead. For now, Glimpse’s release is a research preview centered on visual analysis: inspecting what happened in an image or clip, checking details and explaining the result.