Summary
Elorian has made Glimpse, a visual-reasoning AI model, available in research preview through an API and chat interface. The company reports competitive results on multimodal benchmarks and lists input pricing at $0.60 per million tokens.
Elorian has introduced Glimpse, a visual-reasoning AI model available in research preview through the company's API and interactive chat interface. Elorian says the model can revisit visual representations as it reasons, track objects through motion and changing viewpoints, and connect what it observes to likely physical consequences.
The company presents Glimpse as an early step toward AI systems that understand visual scenes, rather than simply produce descriptions of them. Its examples include following a train as it slows, distinguishing a camera pan from a cut in a video, and predicting how a partly raised tire might fall if its support is lost.
A model designed to revisit what it sees
Elorian says Glimpse can use multimodal tools to write code, transform images and inspect video, and can coordinate subagents that exchange text and visual material. In the chat interface, the company says, users can see a visual scratch space as the model works. Demonstrations show it zooming into image regions, sampling video more densely around an ambiguous moment, and tracking a feature across frames.
These capabilities are relevant to tasks where the order of events or the arrangement of objects matters. In one video example, Glimpse checks intermediate frames to determine whether a camera movement is continuous. In a robotics demonstration, it plans steps for a vision-language-action policy and checks the robot's camera views as the task proceeds. Elorian says both Glimpse and GPT-6 Astra completed the same LIBERO task; it reports a cost of $0.035 for Glimpse's planning in that episode, compared with $0.249 for Astra.
The demonstrations illustrate perception and planning, not direct robot control by Glimpse. Elorian describes direct action outputs, calibrated scene reconstruction and robot control as work still ahead.
Reported benchmark results and pricing
In its published benchmark table, Elorian reports Glimpse scores of 42.70 on WGO Bench for manipulation-event understanding, 83.36 on FSC-147 for open-vocabulary counting, 82.66 on UGC-VideoCap for short-form video captioning and 92.90 on CharXiv-R for scientific chart reasoning. The corresponding GPT-6 Astra scores listed by Elorian are 39.42, 82.02, 80.32 and 92.10. On ERQA, an embodied spatial-reasoning evaluation, both models score 83.75.
These are company-reported results. Elorian says its comparisons generally use each model's highest thinking effort and native agent harness; the configurations include different tools for some models and video evaluations. The table also notes that HSS results include figures reported in a paper, so the scores should be read in light of those evaluation conditions.
Elorian lists Glimpse input tokens at $0.60 per million, uncached, and output tokens at $3.00 per million. The company positions the model as substantially smaller and cheaper than the frontier models in its comparison. Glimpse's research-preview release makes its visual reasoning capabilities available for testing, while the company's stated roadmap points toward more complete scene understanding and robot interaction.