Elorian's Glimpse visual reasoning model arrives on Vercel AI Gateway

Vercel has added Glimpse, a visual reasoning model from Elorian, to its AI Gateway. According to the Vercel blog, developers can now call the model through the gateway's unified API using the identifier elorian/glimpse.

What Glimpse does

Glimpse takes both images and video as input. Vercel describes a model that can track objects as they move, make sense of how elements are positioned relative to one another in a scene, and link an action shown in footage to the physical result it produces.

What distinguishes it, according to the announcement, is that the model can go back to the visual evidence while it works. Instead of a single look and one answer, it can magnify a small detail, review a brief stretch of video again, or follow one feature from frame to frame.

Vercel points to use cases such as analyzing scientific charts, captioning video, counting objects visually, and annotating robotics data. Its examples include judging whether a train in a clip is decelerating, keeping objects separate while the camera itself moves, and marking when an action has been completed in robot footage.

How developers call it

The model is exposed under the name elorian/glimpse. With AI SDK 7, video can be included as a file part in a request; Vercel notes that the placeholder URL in its example should be swapped for a public or signed HTTPS link to the actual video.

Developers can also experiment with the model in Vercel's model playground. For tools such as Claude Code, Codex and Cursor, the gateway can be configured by installing the latest Vercel CLI and running its ai-gateway setup command.

What the gateway layers on top

AI Gateway, as Vercel describes it, provides one API for invoking models, monitoring usage and spend, and setting up retries, failover and performance tuning, which the company says delivers better uptime than calling providers directly. It also includes custom reporting, Zero Data Retention support, budgets tied to API keys, and routing rules.

On pricing, Vercel says the gateway matches what providers charge, applies no markup, and takes no platform fee on inference, including when developers bring their own keys.

Why it matters

Video remains one of the harder inputs for AI systems to reason over, and models built specifically for motion and spatial understanding are still relatively rare. Glimpse's arrival gives developers on Vercel access to that capability class, including tracking objects across frames and reasoning about physical consequences, through infrastructure many already use for text models.

For teams working on robotics annotation, chart analysis or video captioning, the practical benefit is consolidation: one API, one billing path, and built-in retry and failover behavior. Because the gateway adds no fee on top of provider pricing, trying the model carries no extra platform cost. How well it performs on real workloads is the remaining question, and the model playground offers a low-friction way to find out.